Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Removed extra copy for transpose arrays in dot() #1477

Merged
merged 5 commits into from
Jul 17, 2023
Merged

Removed extra copy for transpose arrays in dot() #1477

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
97817ed
Removed extra copy for strided arrays in dot()
antonwolfy Jul 13, 2023
476009e
Merge branch 'master' into strided_dot
antonwolfy Jul 13, 2023
2233689
Added support of strided arrays
antonwolfy Jul 14, 2023
62a287a
Added support of strided out array
antonwolfy Jul 16, 2023
de44929
Fix handling of 1d and 2d arrays
antonwolfy Jul 16, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 8 additions & 13 deletions dpnp/backend/kernels/dpnp_krnl_common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,12 +227,10 @@ DPCTLSyclEventRef dpnp_dot_c(DPCTLSyclQueueRef q_ref,
DPCTLSyclEventRef event_ref = nullptr;
sycl::queue q = *(reinterpret_cast<sycl::queue *>(q_ref));

DPNPC_ptr_adapter<_DataType_input1> input1_ptr(q_ref, input1_in,
input1_size);
DPNPC_ptr_adapter<_DataType_input2> input2_ptr(q_ref, input2_in,
input2_size);
_DataType_input1 *input1 = input1_ptr.get_ptr();
_DataType_input2 *input2 = input2_ptr.get_ptr();
_DataType_input1 *input1 =
static_cast<_DataType_input1 *>(const_cast<void *>(input1_in));
_DataType_input2 *input2 =
static_cast<_DataType_input2 *>(const_cast<void *>(input2_in));
_DataType_output *result = reinterpret_cast<_DataType_output *>(result_out);

if (!input1_size || !input2_size) {
Expand All @@ -257,10 +255,12 @@ DPCTLSyclEventRef dpnp_dot_c(DPCTLSyclQueueRef q_ref,
// if both arrays are vectors
if ((input1_ndim == 1) && (input2_ndim == 1)) {
assert(input1_size == input2_size);

sycl::event event = dot(q, result, input1, input2, input1_strides[0],
input2_strides[0], input1_size);
event.wait();
return event_ref;

event_ref = reinterpret_cast<DPCTLSyclEventRef>(&event);
return DPCTLEvent_Copy(event_ref);
}

// 1D vector
Expand Down Expand Up @@ -318,10 +318,6 @@ DPCTLSyclEventRef dpnp_dot_c(DPCTLSyclQueueRef q_ref,
// (looks like there are such another cases)

if (ext_input1_ndim == 2 && ext_input2_ndim == 2) {
// there is a difference of behavior with trans and sizes params in previous
// version of GEMM only new version is supported, in case of old version
// computation goes in common way
#if INTEL_MKL_VERSION >= 20210004
// is mat1 F-contiguous, C-contiguous
bool mat1_f_contig =
(((ext_input1_shape[0] == 1) || (ext_input1_strides[0] == 1)) &&
Expand Down Expand Up @@ -389,7 +385,6 @@ DPCTLSyclEventRef dpnp_dot_c(DPCTLSyclQueueRef q_ref,
} catch (const std::exception &e) {
// do nothing, proceed to general case
}
#endif
}
}
}
Expand Down
4 changes: 2 additions & 2 deletions dpnp/dpnp_iface_linearalgebra.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,14 +111,14 @@ def dot(x1, x2, out=None, **kwargs):
# TODO: copy_when_strides=False (now it's done for faster implementation with transpose arrays)
x1_desc = dpnp.get_dpnp_descriptor(
x1,
copy_when_strides=True,
copy_when_strides=False,
copy_when_nondefault_queue=False,
alloc_usm_type=usm_type,
alloc_queue=queue,
)
x2_desc = dpnp.get_dpnp_descriptor(
x2,
copy_when_strides=True,
copy_when_strides=False,
copy_when_nondefault_queue=False,
alloc_usm_type=usm_type,
alloc_queue=queue,
Expand Down