+
Skip to content

Properly setting read_dst for gemm in forward/backward pass #777

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
merged 2 commits into from
May 9, 2023
Merged

Properly setting read_dst for gemm in forward/backward pass #777

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3daa2c7
Properly setting read_dst for gemm in forward/backward pass
coreylowman May 9, 2023
18d4c81
Fix clippy warning
coreylowman May 9, 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
3 changes: 3 additions & 0 deletions src/tensor_ops/conv2d/cpu_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,6 +80,7 @@ impl Cpu {
for g in 0..op.groups {
Self::matmul(
(m, k, n),
false,
filters[g * m * k..].as_ptr(),
[k, 1],
buf[g * k * n..].as_ptr(),
Expand Down Expand Up @@ -134,6 +135,7 @@ impl Cpu {
for g in 0..op.groups {
Self::matmul(
(m, k, n),
true,
filters_tr[g * m * k..].as_ptr(),
[k, 1],
buf[g * k * n..].as_ptr(),
Expand All @@ -153,6 +155,7 @@ impl Cpu {
for g in 0..op.groups {
Self::matmul(
(m, k, n),
true,
img[g * m * k..].as_ptr(),
[k, 1],
buf[g * k * n..].as_ptr(),
Expand Down
3 changes: 3 additions & 0 deletions src/tensor_ops/convtrans2d/cpu_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,6 +81,7 @@ impl Cpu {
let n = op.w_out * op.h_out;
Self::matmul(
(m, k, n),
false,
filters.as_ptr(),
[k, 1],
buf.as_ptr(),
Expand Down Expand Up @@ -133,6 +134,7 @@ impl Cpu {
let n = op.h_in * op.w_in;
Self::matmul(
(m, k, n),
true,
filters_tr.as_ptr(),
[k, 1],
buf.as_ptr(),
Expand All @@ -150,6 +152,7 @@ impl Cpu {
let n = op.chan_out * op.kernel * op.kernel;
Self::matmul(
(m, k, n),
true,
img.as_ptr(),
[k, 1],
buf.as_ptr(),
Expand Down
37 changes: 29 additions & 8 deletions src/tensor_ops/matmul/cpu_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,8 +30,10 @@ fn naive_gemm<F: num_traits::Float + std::ops::AddAssign, M: Dim, K: Dim, N: Dim
}

pub(crate) trait MatMulImpl<E> {
#[allow(clippy::too_many_arguments)]
fn matmul<M: Dim, K: Dim, N: Dim>(
dims: (M, K, N),
accum: bool,
ap: *const E,
a_strides: [usize; 2],
bp: *const E,
Expand All @@ -46,6 +48,7 @@ impl MatMulImpl<half::f16> for Cpu {
#[inline]
fn matmul<M: Dim, K: Dim, N: Dim>(
(m, k, n): (M, K, N),
accum: bool,
ap: *const half::f16,
a_strides: [usize; 2],
bp: *const half::f16,
Expand All @@ -65,14 +68,18 @@ impl MatMulImpl<half::f16> for Cpu {
cp as *mut gemm::f16,
c_strides[1] as isize,
c_strides[0] as isize,
true,
ap as *mut gemm::f16,
accum,
ap as *const gemm::f16,
a_strides[1] as isize,
a_strides[0] as isize,
bp as *mut gemm::f16,
bp as *const gemm::f16,
b_strides[1] as isize,
b_strides[0] as isize,
gemm::f16::ONE,
if accum {
gemm::f16::ONE
} else {
gemm::f16::ZERO
},
gemm::f16::ONE,
false,
false,
Expand All @@ -87,6 +94,7 @@ impl MatMulImpl<f32> for Cpu {
#[inline]
fn matmul<M: Dim, K: Dim, N: Dim>(
(m, k, n): (M, K, N),
accum: bool,
ap: *const f32,
a_strides: [usize; 2],
bp: *const f32,
Expand All @@ -106,14 +114,14 @@ impl MatMulImpl<f32> for Cpu {
cp,
c_strides[1] as isize,
c_strides[0] as isize,
true,
accum,
ap,
a_strides[1] as isize,
a_strides[0] as isize,
bp,
b_strides[1] as isize,
b_strides[0] as isize,
1.0,
if accum { 1.0 } else { 0.0 },
1.0,
false,
false,
Expand All @@ -128,6 +136,7 @@ impl MatMulImpl<f64> for Cpu {
#[inline]
fn matmul<M: Dim, K: Dim, N: Dim>(
(m, k, n): (M, K, N),
accum: bool,
ap: *const f64,
a_strides: [usize; 2],
bp: *const f64,
Expand All @@ -147,14 +156,14 @@ impl MatMulImpl<f64> for Cpu {
cp,
c_strides[1] as isize,
c_strides[0] as isize,
true,
accum,
ap,
a_strides[1] as isize,
a_strides[0] as isize,
bp,
b_strides[1] as isize,
b_strides[0] as isize,
1.0,
if accum { 1.0 } else { 0.0 },
1.0,
false,
false,
Expand All @@ -179,6 +188,7 @@ where
let mut out = self.try_zeros_like(&(m, n))?;
Self::matmul(
(m, k, n),
false,
lhs.data.as_ptr(),
lhs.strides,
rhs.data.as_ptr(),
Expand All @@ -201,6 +211,7 @@ where
let strides = (m, n).strides();
Self::matmul(
(m, n, k),
true,
grad_out.as_ptr(),
strides,
rhs.data.as_ptr(),
Expand All @@ -210,6 +221,7 @@ where
);
Self::matmul(
(k, m, n),
true,
lhs.data.as_ptr(),
[lhs.strides[1], lhs.strides[0]],
grad_out.as_ptr(),
Expand Down Expand Up @@ -237,6 +249,7 @@ where
for i in 0..batch.size() {
Self::matmul(
(m, k, n),
false,
lhs.data[i * lhs.strides[0]..].as_ptr(),
[lhs.strides[1], lhs.strides[2]],
rhs.data.as_ptr(),
Expand All @@ -261,6 +274,7 @@ where
for i in 0..batch.size() {
Self::matmul(
(m, n, k),
true,
grad_out[i * strides[0]..].as_ptr(),
[strides[1], strides[2]],
rhs.data.as_ptr(),
Expand All @@ -270,6 +284,7 @@ where
);
Self::matmul(
(k, m, n),
true,
lhs.data[i * lhs.strides[0]..].as_ptr(),
[lhs.strides[2], lhs.strides[1]],
grad_out[i * strides[0]..].as_ptr(),
Expand Down Expand Up @@ -300,6 +315,7 @@ where
for i in 0..b.size() {
Self::matmul(
(m, k, n),
false,
ap[i * lhs.strides[0]..].as_ptr(),
[lhs.strides[1], lhs.strides[2]],
bp[i * rhs.strides[0]..].as_ptr(),
Expand All @@ -324,6 +340,7 @@ where
for i in 0..b.size() {
Self::matmul(
(m, n, k),
true,
grad_out[i * strides[0]..].as_ptr(),
[strides[1], strides[2]],
rhs.data[i * rhs.strides[0]..].as_ptr(),
Expand All @@ -333,6 +350,7 @@ where
);
Self::matmul(
(k, m, n),
true,
lhs.data[i * lhs.strides[0]..].as_ptr(),
[lhs.strides[2], lhs.strides[1]],
grad_out[i * strides[0]..].as_ptr(),
Expand Down Expand Up @@ -362,6 +380,7 @@ where
for j in 0..s.size() {
Self::matmul(
(m, k, n),
false,
lhs.data[i * lhs.strides[0] + j * lhs.strides[1]..].as_ptr(),
[lhs.strides[2], lhs.strides[3]],
rhs.data[i * rhs.strides[0] + j * rhs.strides[1]..].as_ptr(),
Expand All @@ -388,6 +407,7 @@ where
for j in 0..s.size() {
Self::matmul(
(m, n, k),
true,
grad_out[i * strides[0] + j * strides[1]..].as_ptr(),
[strides[2], strides[3]],
rhs.data[i * rhs.strides[0] + j * rhs.strides[1]..].as_ptr(),
Expand All @@ -397,6 +417,7 @@ where
);
Self::matmul(
(k, m, n),
true,
lhs.data[i * lhs.strides[0] + j * lhs.strides[1]..].as_ptr(),
[lhs.strides[3], lhs.strides[2]],
grad_out[i * strides[0] + j * strides[1]..].as_ptr(),
Expand Down
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载