Skip to content

fixes batch matmul for opencl::cpu backend #2059

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

Merged
umar456 merged 2 commits into from
Feb 25, 2018
Merged

fixes batch matmul for opencl::cpu backend #2059

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ac8088e
fixes batch matmul for opencl::cpu backend
syurkevi Feb 21, 2018
2c73c66
fixes issue 1882 for opencl::cpu backend
syurkevi Feb 22, 2018
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
78 changes: 52 additions & 26 deletions src/backend/opencl/cpu/cpu_blas.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,40 +158,66 @@ Array<T> matmul(const Array<T> &lhs, const Array<T> &rhs,
int M = lDims[aRowDim];
int N = rDims[bColDim];
int K = lDims[aColDim];
dim_t d2 = std::max(lDims[2], rDims[2]);
dim_t d3 = std::max(lDims[3], rDims[3]);
dim4 oDims = af::dim4(M, N, d2, d3);

//FIXME: Leaks on errors.
Array<T> out = createValueArray<T>(af::dim4(M, N, 1, 1), scalar<T>(0));
Array<T> out = createValueArray<T>(oDims, scalar<T>(0));
auto alpha = getScale<T, 1>();
auto beta = getScale<T, 0>();

dim4 lStrides = lhs.strides();
dim4 rStrides = rhs.strides();
using BT = typename blas_base<T>::type;
dim4 oStrides = out.strides();

// get host pointers from mapped memory
auto lPtr = lhs.getMappedPtr();
auto rPtr = rhs.getMappedPtr();
auto oPtr = out.getMappedPtr();

if(rDims[bColDim] == 1) {
N = lDims[aColDim];
gemv_func<T>()(
CblasColMajor, lOpts,
lDims[0], lDims[1],
alpha,
(BT*)lPtr.get(), lStrides[1],
(BT*)rPtr.get(), rStrides[0],
beta,
(BT*)oPtr.get(), 1);
} else {
gemm_func<T>()(
CblasColMajor, lOpts, rOpts,
M, N, K,
alpha,
(BT*)lPtr.get(), lStrides[1],
(BT*)rPtr.get(), rStrides[1],
beta,
(BT*)oPtr.get(), out.dims()[0]);
using BT = typename blas_base<T>::type;
using CBT = const typename blas_base<T>::type;

int batchSize = oDims[2] * oDims[3];

bool is_l_d2_batched = (oDims[2] == lDims[2]);
bool is_l_d3_batched = (oDims[3] == lDims[3]);
bool is_r_d2_batched = (oDims[2] == rDims[2]);
bool is_r_d3_batched = (oDims[3] == rDims[3]);

for(int n = 0; n < batchSize; ++n) {
int w = n / rDims[2];
int z = n - w * rDims[2];

int loff = z * (is_l_d2_batched * lStrides[2]) + w * (is_l_d3_batched * lStrides[3]);
int roff = z * (is_r_d2_batched * rStrides[2]) + w * (is_r_d3_batched * rStrides[3]);

// get host pointers from mapped memory
auto lPtr = lhs.getMappedPtr();
auto rPtr = rhs.getMappedPtr();
auto oPtr = out.getMappedPtr();

CBT *lptr = (CBT*)(lPtr.get() + loff);

@umar456 umar456 Feb 23, 2018

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

static_cast

CBT *rptr = (CBT*)(rPtr.get() + roff);
BT *optr = (BT*)(oPtr.get() + z * oStrides[2] + w * oStrides[3]);

if(rDims[bColDim] == 1) {
dim_t incr = (rOpts == CblasNoTrans) ? rStrides[0] : rStrides[1];
N = lDims[aColDim];
gemv_func<T>()(
CblasColMajor, lOpts,
lDims[0], lDims[1],
alpha,
lptr, lStrides[1],
rptr, incr,
beta,
optr, 1);
} else {
gemm_func<T>()(
CblasColMajor, lOpts, rOpts,
M, N, K,
alpha,
lptr, lStrides[1],
rptr, rStrides[1],
beta,
optr, out.dims()[0]);
}
}

return out;
Expand Down