Skip to content

Commit

Permalink
[compute/cker] Optimize BatchMatMul for x86 (#14305)
Browse files Browse the repository at this point in the history
This commit adds an optimized version of the BatchMatMul kernel. The optimization targets the x86 architecture, in all other cases the code is compiled with existing reference kernel.

The new kernel calls the optimized::Gemm function which uses Eigen internally.

Additionally to avoid code duplication a new BatchMatMulParams struct is introduced and reused in both reference and optimized kernels.

ONE-DCO-1.0-Signed-off-by: Tomasz Dolbniak <[email protected]>
  • Loading branch information
tomdol authored Nov 8, 2024
1 parent da3e986 commit 1e09707
Show file tree
Hide file tree
Showing 4 changed files with 202 additions and 65 deletions.
12 changes: 9 additions & 3 deletions compute/cker/include/cker/operation/BatchMatMul.h
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
#include "cker/Types.h"
#include "cker/Shape.h"
#include "cker/Utils.h"
#include "cker/operation/optimized/BatchMatMul.h"
#include "cker/operation/reference/BatchMatMul.h"

#include <vector>
Expand Down Expand Up @@ -77,7 +78,7 @@ class BatchMatMul
}

void operator()(const Shape &lhs_shape, const float *lhs_data, const Shape &rhs_shape,
const float *rhs_data, bool adj_x, bool adj_y, const Shape &output_shape,
const float *rhs_data, bool adj_x, bool adj_y, const Shape & /*output_shape*/,
float *output_data)
{
// Assume lhs and rhs is not constant
Expand All @@ -102,8 +103,13 @@ class BatchMatMul
// Check accumulative dimensions of lhs and rhs of are equal
assert(Shape::ExtendedShape(5, new_rhs_shape).Dims(4) ==
Shape::ExtendedShape(5, new_lhs_shape).Dims(3));
reference::BatchMatMul(new_rhs_shape, new_rhs_data, new_lhs_shape, new_lhs_data, output_shape,
output_data);

const BatchMatMulParams params{new_rhs_shape, new_lhs_shape};
#if defined(CKER_X86_PLATFORM)
optimized::BatchMatMul(params, new_rhs_data, new_lhs_data, output_data);
#else
reference::BatchMatMul(params, new_rhs_data, new_lhs_data, output_data);
#endif
}

private:
Expand Down
98 changes: 98 additions & 0 deletions compute/cker/include/cker/operation/Helper/BatchMatMulParams.h
Original file line number Diff line number Diff line change
@@ -0,0 +1,98 @@
/*
* Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/

#ifndef __NNFW_CKER_HELPER_BATCH_MAT_MUL_PARAMS_H__
#define __NNFW_CKER_HELPER_BATCH_MAT_MUL_PARAMS_H__

#include "cker/Shape.h"

namespace nnfw
{
namespace cker
{
struct BatchMatMulParams
{
BatchMatMulParams(const Shape &lhs_shape, const Shape &rhs_shape)
{
const Shape extended_lhs_shape = Shape::ExtendedShape(5, lhs_shape);
const Shape extended_rhs_shape = Shape::ExtendedShape(5, rhs_shape);

batch_dim0 = broadcast_dim(extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
batch_dim1 = broadcast_dim(extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
batch_dim2 = broadcast_dim(extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));

lhs_ext0 = extent(extended_lhs_shape, 0);
lhs_ext1 = extent(extended_lhs_shape, 1);
lhs_ext2 = extent(extended_lhs_shape, 2);
rhs_ext0 = extent(extended_rhs_shape, 0);
rhs_ext1 = extent(extended_rhs_shape, 1);
rhs_ext2 = extent(extended_rhs_shape, 2);

// Set params for each matrix multiply.
lhs_rows = extended_lhs_shape.Dims(3);
lhs_cols = extended_lhs_shape.Dims(4);
rhs_rows = extended_rhs_shape.Dims(3);
rhs_cols = extended_rhs_shape.Dims(4);
accum_depth = extended_lhs_shape.Dims(4);
}

int batch_dim0;
int batch_dim1;
int batch_dim2;
int lhs_ext0;
int lhs_ext1;
int lhs_ext2;
int rhs_ext0;
int rhs_ext1;
int rhs_ext2;
int lhs_rows;
int lhs_cols;
int rhs_rows;
int rhs_cols;
int accum_depth;

private:
// Determines which dimension is the broadcast dimension.
int32_t broadcast_dim(int32_t lhs_dim, int32_t rhs_dim)
{
if (lhs_dim == rhs_dim)
return lhs_dim;
if (lhs_dim == 1)
return rhs_dim;
assert(rhs_dim == 1);
return lhs_dim;
};

// Computes the "extent" for iterating on this dimension.
// If we are broadcasting, then don't advance (i.e return 0).
int extent(const Shape &shape, int x)
{
if (shape.Dims(x) == 1)
{
return 0;
}
int prod = 1;
for (int i = x + 1; i < shape.DimensionsCount(); ++i)
{
prod *= shape.Dims(i);
}
return prod;
};
};
} // namespace cker
} // namespace nnfw

#endif
75 changes: 75 additions & 0 deletions compute/cker/include/cker/operation/optimized/BatchMatMul.h
Original file line number Diff line number Diff line change
@@ -0,0 +1,75 @@
/*
* Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/

#ifndef __NNFW_CKER_OPTIMIZED_BATCH_MATMUL_H__
#define __NNFW_CKER_OPTIMIZED_BATCH_MATMUL_H__

#include "cker/Shape.h"
#include "cker/operation/Helper/BatchMatMulParams.h"
#include "cker/operation/optimized/Gemm.h"

namespace nnfw
{
namespace cker
{
namespace optimized
{
#if defined(CKER_X86_PLATFORM)

inline void BatchMatMul(const BatchMatMulParams &params, const float *lhs_data,
const float *rhs_data, float *output_data)
{
MatrixParams<float> lhs_params;
lhs_params.order = Order::kRowMajor;
lhs_params.rows = params.lhs_rows;
lhs_params.cols = params.lhs_cols;

MatrixParams<float> rhs_params;
rhs_params.order = Order::kRowMajor;
rhs_params.rows = params.rhs_rows;
rhs_params.cols = params.rhs_cols;

MatrixParams<float> dst_params;
dst_params.order = Order::kRowMajor;
dst_params.rows = params.lhs_rows;
dst_params.cols = params.rhs_cols;

for (int b0 = 0; b0 < params.batch_dim0; ++b0)
{
for (int b1 = 0; b1 < params.batch_dim1; ++b1)
{
for (int b2 = 0; b2 < params.batch_dim2; ++b2)
{
const float *lhs_ptr =
lhs_data + b0 * params.lhs_ext0 + b1 * params.lhs_ext1 + b2 * params.lhs_ext2;
const float *rhs_ptr =
rhs_data + b0 * params.rhs_ext0 + b1 * params.rhs_ext1 + b2 * params.rhs_ext2;
float *out_ptr = output_data + ((b0 * params.batch_dim1 * params.batch_dim2) +
b1 * params.batch_dim2 + b2) *
params.lhs_rows * params.rhs_cols;

optimized::Gemm(lhs_params, lhs_ptr, rhs_params, rhs_ptr, dst_params, out_ptr,
GemmParams<float, float>{});
}
}
}
}
#endif
} // namespace optimized
} // namespace cker
} // namespace nnfw

#endif // __NNFW_CKER_OPTIMIZED_BATCH_MATMUL_H__
82 changes: 20 additions & 62 deletions compute/cker/include/cker/operation/reference/BatchMatMul.h
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@

#include "cker/Types.h"
#include "cker/Shape.h"
#include "cker/operation/Helper/BatchMatMulParams.h"

namespace nnfw
{
Expand All @@ -28,77 +29,34 @@ namespace cker
namespace reference
{

inline void BatchMatMul(const Shape &lhs_shape, const float *lhs_data, const Shape &rhs_shape,
const float *rhs_data, const Shape &, float *output_data)
inline void BatchMatMul(const BatchMatMulParams &params, const float *lhs_data,
const float *rhs_data, float *output_data)
{
const Shape extended_lhs_shape = Shape::ExtendedShape(5, lhs_shape);
const Shape extended_rhs_shape = Shape::ExtendedShape(5, rhs_shape);

// Determine which dimension is the broadcast dimension.
auto broadcast_dim = [](int lhs_dim, int rhs_dim) {
if (lhs_dim == rhs_dim)
return lhs_dim;
if (lhs_dim == 1)
return rhs_dim;
assert(rhs_dim == 1);
return lhs_dim;
};

// Compute the "extent" for iterating on this dimension.
// If we are broadcasting, then don't advance (i.e return 0).
auto extent = [](const Shape &shape, int x) {
if (shape.Dims(x) == 1)
{
return 0;
}
int prod = 1;
for (int i = x + 1; i < shape.DimensionsCount(); ++i)
{
prod *= shape.Dims(i);
}
return prod;
};

const int batch_dim0 = broadcast_dim(extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
const int batch_dim1 = broadcast_dim(extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
const int batch_dim2 = broadcast_dim(extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));

const int lhs_ext0 = extent(extended_lhs_shape, 0);
const int lhs_ext1 = extent(extended_lhs_shape, 1);
const int lhs_ext2 = extent(extended_lhs_shape, 2);
const int rhs_ext0 = extent(extended_rhs_shape, 0);
const int rhs_ext1 = extent(extended_rhs_shape, 1);
const int rhs_ext2 = extent(extended_rhs_shape, 2);

// Set params for each matrix multiply.
const int lhs_rows = extended_lhs_shape.Dims(3);
const int rhs_cols = extended_rhs_shape.Dims(4);
const int accum_depth = extended_lhs_shape.Dims(4);

for (int b0 = 0; b0 < batch_dim0; ++b0)
for (int b0 = 0; b0 < params.batch_dim0; ++b0)
{
const float *lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
const float *rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
for (int b1 = 0; b1 < batch_dim1; ++b1)
const float *lhs_ptr0 = lhs_data + (b0 * params.lhs_ext0);
const float *rhs_ptr0 = rhs_data + (b0 * params.rhs_ext0);
for (int b1 = 0; b1 < params.batch_dim1; ++b1)
{
const float *lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
const float *rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
for (int b2 = 0; b2 < batch_dim2; ++b2)
const float *lhs_ptr1 = lhs_ptr0 + b1 * params.lhs_ext1;
const float *rhs_ptr1 = rhs_ptr0 + b1 * params.rhs_ext1;
for (int b2 = 0; b2 < params.batch_dim2; ++b2)
{
const float *lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
const float *rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
float *out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) + b1 * batch_dim2 + b2) *
lhs_rows * rhs_cols;
for (int j = 0; j < rhs_cols; ++j)
const float *lhs_ptr2 = lhs_ptr1 + b2 * params.lhs_ext2;
const float *rhs_ptr2 = rhs_ptr1 + b2 * params.rhs_ext2;
float *out_ptr = output_data + ((b0 * params.batch_dim1 * params.batch_dim2) +
b1 * params.batch_dim2 + b2) *
params.lhs_rows * params.rhs_cols;
for (int j = 0; j < params.rhs_cols; ++j)
{
for (int i = 0; i < lhs_rows; ++i)
for (int i = 0; i < params.lhs_rows; ++i)
{
float total = 0.f;
for (int k = 0; k < accum_depth; ++k)
for (int k = 0; k < params.accum_depth; ++k)
{
total += lhs_ptr2[accum_depth * i + k] * rhs_ptr2[j * accum_depth + k];
total += lhs_ptr2[params.accum_depth * i + k] * rhs_ptr2[j * params.accum_depth + k];
}
int idx = lhs_rows * j + i;
int idx = params.lhs_rows * j + i;
out_ptr[idx] = total;
}
}
Expand Down

0 comments on commit 1e09707

Please sign in to comment.