未验证 提交 412ee03a 编写于 作者: S smilejames 提交者: GitHub

Merge pull request #919 from smilejames/develop

add border check in gemm
......@@ -3002,14 +3002,21 @@ void Sgemm(int m, int n, int k, float alpha, const float *A, int lda,
NC = L2 / (KC * sizeof(float));
// make sure MC is multiple of MR, and NC is multiple of NR
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC = NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA = static_cast<float *>(
......@@ -3067,14 +3074,21 @@ void SgemmWithBn(int m, int n, int k, float alpha, const float *A, int lda,
NC = L2 / (KC * sizeof(float));
// make sure MC is multiple of MR, and NC is multiple of NR
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC = NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA = static_cast<float *>(
......@@ -3133,14 +3147,21 @@ void SgemmWithPRelu(int m, int n, int k, const float *A, int lda,
NC = L2 / (KC * sizeof(float));
// make sure MC is multiple of MR, and NC is multiple of NR
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC = NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA = static_cast<float *>(
......@@ -3203,9 +3224,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
if (m > n) {
// 对 A 分块
MC = L1 / (KC * sizeof(float));
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// 补齐 B
NC = (n + NR - 1) / NR * NR;
......@@ -3227,9 +3252,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
} else {
// 对 B 分块
NC = L1 / (KC * sizeof(float));
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC = NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// 补齐 A
MC = (m + MR - 1) / MR * MR;
......@@ -3311,9 +3340,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
if (m > n) {
// 对 A 分块
MC = L1 / (KC * sizeof(float));
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// 补齐 B
NC = (n + NR - 1) / NR * NR;
......@@ -3335,9 +3368,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
} else {
// 对 B 分块
NC = L1 / (KC * sizeof(float));
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC == NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// 补齐 A
MC = (m + MR - 1) / MR * MR;
......@@ -3430,9 +3467,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
if (m > n) {
// 对 A 分块
MC = L1 / (KC * sizeof(float));
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
if (MC == 0) {
MC = MR;
} else {
int mblock_num = (m + MC - 1) / MC;
MC = (m + mblock_num - 1) / mblock_num;
MC = (MC + MR - 1) / MR * MR;
}
// 补齐 B
NC = (n + NR - 1) / NR * NR;
......@@ -3454,9 +3495,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
} else {
// 对 B 分块
NC = L1 / (KC * sizeof(float));
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
if (NC == 0) {
NC = NR;
} else {
int nblock_num = (n + NC - 1) / NC;
NC = (n + nblock_num - 1) / nblock_num;
NC = (NC + NR - 1) / NR * NR;
}
// 补齐 A
MC = (m + MR - 1) / MR * MR;
......
......@@ -28,16 +28,9 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
auto dim_a = matrix_a.dims();
auto dim_b = matrix_b.dims();
auto dim_out = matrix_out->dims();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE(
dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
int M = dim_out[0];
int N = dim_out[1];
......@@ -61,16 +54,9 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
auto dim_a = matrix_a.dims();
auto dim_b = matrix_b.dims();
auto dim_out = matrix_out->dims();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE(
dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
int M = dim_out[0];
int N = dim_out[1];
......@@ -95,16 +81,9 @@ void matmulWithPRelu(const framework::Tensor &matrix_a, bool trans_a,
auto dim_a = matrix_a.dims();
auto dim_b = matrix_b.dims();
auto dim_out = matrix_out->dims();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE(
dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
int M = dim_out[0];
int N = dim_out[1];
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册