Skip to content

Commit 5c6c720

Browse files
AlexanderKalistratovAlexanderKalistratov
committed
Initial implementation of Deformable Convolution workload
1 parent 09dc9ed commit 5c6c720

11 files changed

Lines changed: 786 additions & 0 deletions

File tree

‎dpbench/benchmarks/CMakeLists.txt‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,7 @@ add_subdirectory(rambo)
55
add_subdirectory(kmeans)
66
add_subdirectory(knn)
77
add_subdirectory(gpairs)
8+
add_subdirectory(deformable_convolution)
89
add_subdirectory(dbscan)
910

1011
# generate dpcpp version into json

‎dpbench/benchmarks/deformable_convolution/CMakeLists.txt‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
add_subdirectory(deformable_convolution_sycl_native_ext)

‎dpbench/benchmarks/deformable_convolution/__init__.py‎

Lines changed: 37 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,37 @@
1+
# Copyright 2022 Intel Corporation
2+
#
3+
# SPDX-License-Identifier: Apache 2.0
4+
5+
from .deformable_convolution_initialize import initialize
6+
from .deformable_convolution_numba_npr import deformable_convolution as deformable_convolution_numba_npr
7+
from .deformable_convolution_numpy import deformable_convolution as deformable_convolution_numpy
8+
from .deformable_convolution_sycl_native_ext import deformable_convolution_sycl
9+
10+
__all__ = [
11+
"initialize",
12+
"deformable_convolution_numba_npr",
13+
"deformable_convolution_numba_numpy",
14+
"deformable_convolution_sycl",
15+
]
16+
17+
"""l2-norm calculation of n vectors
18+
19+
Input
20+
---------
21+
npoints: int
22+
number of vectors
23+
dims: int
24+
dimension of single vector
25+
seed: int
26+
random seed to generate random number
27+
28+
Output
29+
-------
30+
d: array
31+
l2 norm of each vector
32+
33+
Method
34+
------
35+
||Vj||2=sqrt(sum(Xj[i]*Xj[i]))
36+
here i is 0->dims, j is 0->npoints
37+
"""

‎dpbench/benchmarks/deformable_convolution/deformable_convolution_initialize.py‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
# Copyright 2022 Intel Corp.
2+
#
3+
# SPDX-License-Identifier: Apache-2.0
4+
5+
6+
def initialize(batch,
7+
in_channels, in_height, in_width,
8+
out_channels, out_height, out_width,
9+
kernel_height, kernel_width,
10+
stride_y, stride_x,
11+
dilation_y, dilation_x,
12+
pad_y, pad_x,
13+
groups,
14+
deformable_groups,
15+
dtype,
16+
seed):
17+
import numpy as np
18+
import numpy.random as default_rng
19+
default_rng.seed(seed)
20+
21+
return (
22+
default_rng.random((batch, in_channels, in_height, in_width)).astype(dtype),
23+
# np.ones((batch, in_channels, in_height, in_width)).astype(dtype),
24+
np.zeros((batch, out_channels, out_height, out_width)).astype(dtype),
25+
# np.zeros((kernel_height, kernel_width, 2, out_height, out_width)).astype(dtype),
26+
2*default_rng.random((kernel_height, kernel_width, 2, out_height, out_width)).astype(dtype) - 1,
27+
# default_rng.random((out_channels, in_channels, kernel_height, kernel_width)).astype(dtype),
28+
np.ones((out_channels, in_channels, kernel_height, kernel_width)).astype(dtype),
29+
default_rng.random(out_channels).astype(dtype),
30+
# np.ones((out_channels,)).astype(dtype),
31+
np.zeros((in_channels, kernel_height, kernel_width, out_height, out_width)).astype(dtype)
32+
)

‎dpbench/benchmarks/deformable_convolution/deformable_convolution_numba_npr.py‎

Lines changed: 81 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,81 @@
1+
# Copyright 2022 Intel Corp.
2+
#
3+
# SPDX-License-Identifier: Apache-2.0
4+
5+
import math
6+
import numpy as np
7+
from numba import njit, prange
8+
9+
@njit(parallel=True, inline='always', fastmath=True)
10+
def bilinear(input, offset_y, offset_x):
11+
height, width = input.shape
12+
start_x = int(math.floor(offset_x))
13+
start_x_weight = 1 - (offset_x - start_x)
14+
start_y = int(math.floor(offset_y))
15+
start_y_weight = 1 - (offset_y - start_y)
16+
17+
output = 0
18+
if offset_x >= width or offset_y >= height or offset_x <= -1 or offset_y <= -1:
19+
return output
20+
21+
if start_x >= 0 and start_y >= 0:
22+
w = start_x_weight*start_y_weight
23+
output += w*input[start_y, start_x]
24+
25+
if start_x + 1 < width and start_y >= 0:
26+
w = (1 - start_x_weight)*start_y_weight
27+
output += w*input[start_y, start_x + 1]
28+
29+
if start_x >=0 and start_y + 1 < height:
30+
w = start_x_weight*(1 - start_y_weight)
31+
output += w*input[start_y + 1, start_x]
32+
33+
if start_x + 1 < width and start_y + 1 < height:
34+
w = (1 - start_x_weight)*(1 - start_y_weight)
35+
output += w*input[start_y + 1, start_x + 1]
36+
37+
return output
38+
39+
@njit(parallel=True, fastmath=True)
40+
def deform(input, offset, output, stride, pad, dilation, groups, deformable_groups):
41+
k_height, k_width, _, out_height, out_width = offset.shape
42+
channels, _, _ = input.shape
43+
44+
k_h_m = (k_height - 1)//2
45+
k_w_m = (k_width - 1)//2
46+
47+
for ckhkw in prange(channels*k_height*k_width):
48+
for h in prange(out_height):
49+
for w in prange(out_width):
50+
c = ckhkw//(k_height*k_width)
51+
khkw = ckhkw%(k_height*k_width)
52+
kh = khkw//k_width
53+
kw = khkw%k_width
54+
55+
offset_y = offset[kh, kw, 1, h, w] + h*stride[0] + (kh - k_h_m)*dilation[0] - (pad[0] - k_h_m)
56+
offset_x = offset[kh, kw, 0, h, w] + w*stride[1] + (kw - k_w_m)*dilation[1] - (pad[1] - k_w_m)
57+
58+
output[c, kh, kw, h, w] = bilinear(input[c], offset_y, offset_x)
59+
60+
@njit(parallel=True, fastmath=True)
61+
def deformable_convolution_b1(input, output, offset, weights, bias, tmp, stride, pad, dilation, groups, deformable_groups):
62+
out_channels, height, width = output.shape
63+
_, in_channels, k_height, k_width = weights.shape
64+
65+
deform(input, offset, tmp, stride, pad, dilation, groups, deformable_groups)
66+
67+
tmp = tmp.reshape((in_channels*k_height*k_width, height*width))
68+
69+
_weights = weights.reshape((out_channels, in_channels*k_height*k_width))
70+
_output = output.reshape((out_channels, height*width))
71+
np.dot(_weights, tmp, _output)
72+
73+
_bias = bias.reshape((out_channels, 1))
74+
_output[:] = _output + _bias
75+
76+
@njit(parallel=True)
77+
def deformable_convolution(input, output, offset, weights, bias, tmp, stride_y, stride_x, pad_y, pad_x, dilation_y, dilation_x, groups, deformable_groups):
78+
batch, _, _, _ = input.shape
79+
for b in range(batch):
80+
deformable_convolution_b1(input[b], output[b], offset, weights, bias, tmp, (stride_y, stride_x), (pad_y, pad_x), (dilation_y, dilation_x), groups, deformable_groups)
81+

‎dpbench/benchmarks/deformable_convolution/deformable_convolution_numpy.py‎

Lines changed: 81 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,81 @@
1+
# Copyright 2022 Intel Corp.
2+
#
3+
# SPDX-License-Identifier: Apache-2.0
4+
5+
import math
6+
import numpy as np
7+
from numba import njit, prange
8+
9+
@njit(parallel=True, inline='always', fastmath=True)
10+
def bilinear(input, offset_y, offset_x):
11+
height, width = input.shape
12+
start_x = int(math.floor(offset_x))
13+
start_x_weight = 1 - (offset_x - start_x)
14+
start_y = int(math.floor(offset_y))
15+
start_y_weight = 1 - (offset_y - start_y)
16+
17+
output = 0
18+
if start_x >= width or start_y >= height or start_x < -1 or start_y < -1:
19+
return output
20+
21+
if start_x >= 0 and start_y >= 0:
22+
w = start_x_weight*start_y_weight
23+
output += w*input[start_y, start_x]
24+
25+
if start_x + 1 < width and start_y >= 0:
26+
w = (1 - start_x_weight)*start_y_weight
27+
output += w*input[start_y, start_x + 1]
28+
29+
if start_x >=0 and start_y + 1 < height:
30+
w = start_x_weight*(1 - start_y_weight)
31+
output += w*input[start_y + 1, start_x]
32+
33+
if start_x + 1 < width and start_y + 1 < height:
34+
w = (1 - start_x_weight)*(1 - start_y_weight)
35+
output += w*input[start_y + 1, start_x + 1]
36+
37+
return output
38+
39+
@njit(parallel=True, fastmath=True)
40+
def deform(input, offset, output, stride, pad, dilation, groups, deformable_groups):
41+
k_height, k_width, _, out_height, out_width = offset.shape
42+
channels, _, _ = input.shape
43+
44+
k_h_m = (k_height - 1)//2
45+
k_w_m = (k_width - 1)//2
46+
47+
for ckhkw in prange(channels*k_height*k_width):
48+
for h in prange(out_height):
49+
for w in prange(out_width):
50+
c = ckhkw//(k_height*k_width)
51+
khkw = ckhkw%(k_height*k_width)
52+
kh = khkw//k_width
53+
kw = khkw%k_width
54+
55+
offset_y = offset[kh, kw, 1, h, w] + h*stride[0] + (kh - k_h_m)*dilation[0] - (pad[0] - k_h_m)
56+
offset_x = offset[kh, kw, 0, h, w] + w*stride[1] + (kw - k_w_m)*dilation[1] - (pad[1] - k_w_m)
57+
58+
output[c, kh, kw, h, w] = bilinear(input[c], offset_y, offset_x)
59+
60+
# @njit(parallel=True, fastmath=True)
61+
def deformable_convolution_b1(input, output, offset, weights, bias, tmp, stride, pad, dilation, groups, deformable_groups):
62+
out_channels, height, width = output.shape
63+
_, in_channels, k_height, k_width = weights.shape
64+
65+
deform(input, offset, tmp, stride, pad, dilation, groups, deformable_groups)
66+
67+
tmp = tmp.reshape((in_channels*k_height*k_width, height*width))
68+
69+
_weights = weights.reshape((out_channels, in_channels*k_height*k_width))
70+
_output = output.reshape((out_channels, height*width))
71+
np.dot(_weights, tmp, _output)
72+
73+
_bias = bias.reshape((out_channels, 1))
74+
_output[:] = _output + _bias
75+
76+
# @njit(parallel=True)
77+
def deformable_convolution(input, output, offset, weights, bias, tmp, stride_y, stride_x, pad_y, pad_x, dilation_y, dilation_x, groups, deformable_groups):
78+
batch, _, _, _ = input.shape
79+
for b in range(batch):
80+
deformable_convolution_b1(input[b], output[b], offset, weights, bias, tmp, (stride_y, stride_x), (pad_y, pad_x), (dilation_y, dilation_x), groups, deformable_groups)
81+

‎dpbench/benchmarks/deformable_convolution/deformable_convolution_sycl_native_ext/CMakeLists.txt‎

Lines changed: 20 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,20 @@
1+
set(py_module_name _deformable_convolution_sycl)
2+
pybind11_add_module(${py_module_name}
3+
MODULE
4+
deformable_convolution_sycl/impl.cpp
5+
)
6+
find_package(MKL CONFIG REQUIRED)
7+
8+
# target_include_directories(${py_module_name} PUBLIC ${Dpctl_INCLUDE_DIRS} $)
9+
target_compile_options(${py_module_name} PUBLIC $<TARGET_PROPERTY:MKL::MKL,INTERFACE_COMPILE_OPTIONS>)
10+
target_include_directories(${py_module_name} PUBLIC ${Dpctl_INCLUDE_DIRS} $<TARGET_PROPERTY:MKL::MKL,INTERFACE_INCLUDE_DIRECTORIES>)
11+
target_link_libraries(${py_module_name} PUBLIC $<LINK_ONLY:MKL::MKL> ${MKL_SYCL})
12+
13+
message(STATUS "${MKL_IMPORTED_TARGETS}")
14+
message(STATUS "${INTERFACE_COMPILE_OPTIONS}")
15+
message(STATUS "${INTERFACE_INCLUDE_DIRECTORIES}")
16+
17+
file(RELATIVE_PATH py_module_dest ${CMAKE_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR})
18+
install(TARGETS ${py_module_name}
19+
DESTINATION ${py_module_dest}/deformable_convolution_sycl
20+
)

‎dpbench/benchmarks/deformable_convolution/deformable_convolution_sycl_native_ext/__init__.py‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,7 @@
1+
# Copyright 2022 Intel Corporation
2+
#
3+
# SPDX-License-Identifier: Apache 2.0
4+
5+
from .deformable_convolution_sycl._deformable_convolution_sycl import deformable_convolution as deformable_convolution_sycl
6+
7+
__all__ = ["deformable_convolution_sycl"]

0 commit comments

Comments
 (0)