Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 b35e89d6 编写于 作者: W Wilber 提交者： Xiaoyang LI
浏览文件
操作

Add matmul op (#1837) 

* test=develop add matmul_op

* use lite::arm::math::sgemm func to implement matmul

* test=develop  pre-commit command to run clang-format

* Revert "test=develop  pre-commit command to run clang-format"

This reverts commit 3f56474f.

* test=develop pre-commit command to run clang-format
上级 118ad09e
显示空白变更内容
内联 并排
Showing with 346 addition and 219 deletion
lite/api/_paddle_use_kernels.h
浏览文件 @ b35e89d6
......@@ -35,7 +35,7 @@ USE_LITE_KERNEL(reshape2, kHost, kAny, kAny, def);
#ifdef LITE_WITH_ARM
USE_LITE_KERNEL(fc, kARM, kFloat, kNCHW, def);
USE_LITE_KERNEL(mul, kARM, kFloat, kNCHW, def);
USE_LITE_KERNEL(matmul, kARM, kFloat, kNCHW, def); // for x2paddle
USE_LITE_KERNEL(matmul, kARM, kFloat, kNCHW, def);
USE_LITE_KERNEL(scale, kARM, kFloat, kNCHW, def);
USE_LITE_KERNEL(softmax, kARM, kFloat, kNCHW, def);
USE_LITE_KERNEL(lrn, kARM, kFloat, kNCHW, def);
......
lite/api/_paddle_use_ops.h
浏览文件 @ b35e89d6
......@@ -19,7 +19,7 @@
#include "paddle_lite_factory_helper.h" // NOLINT
USE_LITE_OP(mul);
USE_LITE_OP(matmul); // for x2paddle
USE_LITE_OP(matmul);
USE_LITE_OP(fc);
USE_LITE_OP(relu);
USE_LITE_OP(relu6);
......
lite/kernels/arm/matmul_compute.cc
浏览文件 @ b35e89d6
......@@ -36,6 +36,7 @@ void MatMulCompute::Run() {
auto x_dims = param.X->dims();
auto y_dims = param.Y->dims();
auto o_dims = param.Out->dims();
bool x_transpose = param.transpose_X;
bool y_transpose = param.transpose_Y;
float alpha = param.alpha;
......@@ -44,136 +45,102 @@ void MatMulCompute::Run() {
if (x_dims.size() > 2 && y_dims.size() >= 2) {
// x: [B, ..., M, K], y: [B, ..., K, N], out: [B, ..., M, N]
// x: [B, M, K], y: [K, N], out: [B, M, N]
if (x_transpose || y_transpose) {
LOG(FATAL) << "not supported transpose for x or y.";
}
if (!x_transpose && !y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 1], y_dims[y_dims.size() - 2])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ")";
<< ") x_transpose is " << x_transpose << "y_transpose is "
<< y_transpose;
} else if (!x_transpose && y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 1], y_dims[y_dims.size() - 1])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ") x_transpose is " << x_transpose << "y_transpose is "
<< y_transpose;
} else if (x_transpose && !y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 2], y_dims[y_dims.size() - 2])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ") x_transpose is " << x_transpose << "y_transpose is "
<< y_transpose;
} else {
CHECK_EQ(x_dims[x_dims.size() - 2], y_dims[y_dims.size() - 1])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ") x_transpose is " << x_transpose << "y_transpose is "
<< y_transpose;
}
if (y_dims.size() > 2) {
int lda, ldb, ldc;
if (!x_transpose) {
m_ = x_dims[x_dims.size() - 2];
k_ = y_dims[y_dims.size() - 2];
k_ = x_dims[x_dims.size() - 1];
lda = k_;
} else {
m_ = x_dims[x_dims.size() - 1];
k_ = x_dims[x_dims.size() - 2];
lda = m_;
}
if (!y_transpose) {
n_ = y_dims[y_dims.size() - 1];
int hblock = lite::arm::math::get_hblock(ctx.arch());
int m_round = 0;
m_round = hblock * ((m_ + hblock - 1) / hblock);
ctx.ExtendWorkspace(m_round * k_ * sizeof(float));
ldb = n_;
} else {
n_ = y_dims[y_dims.size() - 2];
ldb = k_;
}
ldc = n_;
int x_inner = x_dims[x_dims.size() - 2] * x_dims[x_dims.size() - 1];
int y_inner = y_dims[y_dims.size() - 2] * y_dims[y_dims.size() - 1];
int out_inner = x_dims[x_dims.size() - 2] * y_dims[y_dims.size() - 1];
if (n_ == 1) {
for (size_t i = 0; i < x_dims.count(0, x_dims.size() - 2); ++i) {
lite::arm::math::sgemv(x_data + i * x_inner,
y_data + i * y_inner,
o_data + i * out_inner,
false,
m_,
k_,
false,
nullptr,
false);
}
if (fabsf(alpha - 1.f) > 1e-8f) {
for (size_t i = 0; i < param.Out->dims().production(); ++i) {
o_data[i] *= alpha;
}
int out_inner = o_dims[o_dims.size() - 2] * o_dims[o_dims.size() - 1];
float* x_data_trans = nullptr;
if (x_transpose) {
x_data_trans = static_cast<float*>(malloc(sizeof(float) * x_inner));
}
} else {
if (y_dims.size() > 2) {
for (size_t i = 0; i < x_dims.count(0, x_dims.size() - 2); ++i) {
float* packed_x = static_cast<float*>(ctx.workspace_data<float>()) +
ctx.llc_size() / sizeof(float);
lite::arm::math::prepackA(packed_x,
x_data + i * x_inner,
alpha,
k_,
0,
m_,
0,
k_,
false,
&ctx);
int ldb = n_;
if (y_transpose) {
ldb = k_;
}
lite::arm::math::sgemm_prepack(y_transpose,
lite::arm::math::sgemm(x_transpose,
y_transpose,
m_,
n_,
k_,
packed_x,
alpha,
x_data + i * x_inner,
lda,
y_data + i * y_inner,
ldb,
0.f,
o_data + i * out_inner,
n_,
ldc,
nullptr,
false,
false,
&ctx);
}
}
} else {
m_ = x_dims[x_dims.size() - 2];
k_ = y_dims[0];
n_ = y_dims[1];
int hblock = lite::arm::math::get_hblock(ctx.arch());
int m_round = 0;
m_round = hblock * ((m_ + hblock - 1) / hblock);
ctx.ExtendWorkspace(m_round * k_ * sizeof(float));
int x_inner = x_dims[x_dims.size() - 2] * x_dims[x_dims.size() - 1];
int out_inner = x_dims[x_dims.size() - 2] * y_dims[1];
if (n_ == 1) {
for (size_t i = 0; i < x_dims.count(0, x_dims.size() - 2); ++i) {
lite::arm::math::sgemv(x_data + i * x_inner,
y_data,
o_data + i * out_inner,
false,
m_,
k_,
false,
nullptr,
false);
}
if (fabsf(param.alpha - 1.f) > 1e-8f) {
for (size_t i = 0; i < param.Out->dims().production(); ++i) {
o_data[i] *= param.alpha;
}
}
} else {
for (size_t i = 0; i < x_dims.count(0, x_dims.size() - 2); ++i) {
float* packed_x = static_cast<float*>(ctx.workspace_data<float>()) +
ctx.llc_size() / sizeof(float);
lite::arm::math::prepackA(packed_x,
x_data + i * x_inner,
alpha,
k_,
0,
m_,
0,
k_,
false,
&ctx);
int ldb = n_;
if (y_transpose) {
ldb = k_;
}
lite::arm::math::sgemm_prepack(y_transpose,
lite::arm::math::sgemm(x_transpose,
y_transpose,
m_,
n_,
k_,
packed_x,
alpha,
x_data + i * x_inner,
lda,
y_data,
ldb,
0.f,
o_data + i * out_inner,
n_,
ldc,
nullptr,
false,
false,
&ctx);
}
}
if (x_data_trans) {
free(x_data_trans);
}
} else if (x_dims.size() == 2 && y_dims.size() == 2) {
// x: [M, K], y: [K, N], out: [M, N]
......@@ -198,50 +165,43 @@ void MatMulCompute::Run() {
<< "), x_transpose is " << x_transpose << ", y_transpose is "
<< y_transpose;
}
// not supported transpose
if (x_transpose || y_transpose) {
LOG(FATAL) << "not supported transpose for x and y.";
}
int lda, ldb, ldc;
if (!x_transpose) {
m_ = x_dims[0];
k_ = x_dims[1];
n_ = y_dims[1];
int hblock = lite::arm::math::get_hblock(ctx.arch());
int m_round = 0;
m_round = hblock * ((m_ + hblock - 1) / hblock);
ctx.ExtendWorkspace(m_round * k_ * sizeof(float));
if (n_ == 1) {
lite::arm::math::sgemv(
x_data, y_data, o_data, x_transpose, m_, k_, false, nullptr, false);
if (fabsf(param.alpha - 1.f) > 1e-8f) {
for (size_t i = 0; i < param.Out->dims().production(); ++i) {
o_data[i] *= param.alpha;
}
lda = k_;
} else {
m_ = x_dims[1];
k_ = x_dims[0];
lda = m_;
}
if (!y_transpose) {
n_ = y_dims[1];
ldb = n_;
} else {
float* packed_x = static_cast<float*>(ctx.workspace_data<float>()) +
ctx.llc_size() / sizeof(float);
lite::arm::math::prepackA(
packed_x, x_data, alpha, k_, 0, m_, 0, k_, x_transpose, &ctx);
int ldb = n_;
if (y_transpose) {
n_ = y_dims[0];
ldb = k_;
}
lite::arm::math::sgemm_prepack(y_transpose,
ldc = n_;
lite::arm::math::sgemm(x_transpose,
y_transpose,
m_,
n_,
k_,
packed_x,
alpha,
x_data,
lda,
y_data,
ldb,
0.f,
o_data,
n_,
ldc,
nullptr,
false,
false,
&ctx);
}
} else if (x_dims.size() > 2 && y_dims.size() == 1) {
// x: [B, M, K], y: [K], out: [B, M]
CHECK_EQ(x_dims[x_dims.size() - 1], y_dims[0])
......@@ -267,6 +227,9 @@ void MatMulCompute::Run() {
m_ = x_dims[0];
k_ = 1;
n_ = y_dims[0];
int lda = k_;
int ldb = n_;
int ldc = n_;
if (n_ == 1) {
lite::arm::math::sgemv(
x_data, y_data, o_data, false, m_, k_, false, nullptr, false);
......@@ -276,21 +239,19 @@ void MatMulCompute::Run() {
}
}
} else {
float* packed_x = static_cast<float*>(ctx.workspace_data<float>()) +
ctx.llc_size() / sizeof(float);
lite::arm::math::prepackA(
packed_x, x_data, alpha, k_, 0, m_, 0, k_, false, &ctx);
int ldb = n_;
lite::arm::math::sgemm_prepack(false,
lite::arm::math::sgemm(false,
false,
m_,
n_,
k_,
packed_x,
alpha,
x_data,
lda,
y_data,
ldb,
0.f,
o_data,
n_,
ldc,
nullptr,
false,
false,
......
lite/operators/matmul_op.cc
浏览文件 @ b35e89d6
......@@ -37,14 +37,41 @@ bool MatMulOpLite::InferShape() const {
if (x_dims.size() > 2 && y_dims.size() >= 2) {
// x: [B, ..., M, K], y: [B, ..., K, N], out: [B, ..., M, N]
// x: [B, M, K], y: [K, N], out: [B, M, N]
if (!x_transpose && !y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 1], y_dims[y_dims.size() - 2])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ")";
} else if (!x_transpose && y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 1], y_dims[y_dims.size() - 1])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ")";
} else if (x_transpose && !y_transpose) {
CHECK_EQ(x_dims[x_dims.size() - 2], y_dims[y_dims.size() - 2])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ")";
} else {
CHECK_EQ(x_dims[x_dims.size() - 2], y_dims[y_dims.size() - 1])
<< "not supported x_dims(" << x_dims << ") and y_dims(" << y_dims
<< ")";
}
dim_out_vec.resize(x_dims.size());
for (size_t i = 0; i < x_dims.size() - 1; ++i) {
for (size_t i = 0; i < x_dims.size() - 2; ++i) {
dim_out_vec[i] = x_dims[i];
}
if (!x_transpose && !y_transpose) {
dim_out_vec[x_dims.size() - 2] = x_dims[x_dims.size() - 2];
dim_out_vec[x_dims.size() - 1] = y_dims[y_dims.size() - 1];
} else if (!x_transpose && y_transpose) {
dim_out_vec[x_dims.size() - 2] = x_dims[x_dims.size() - 2];
dim_out_vec[x_dims.size() - 1] = y_dims[y_dims.size() - 2];
} else if (x_transpose && !y_transpose) {
dim_out_vec[x_dims.size() - 2] = x_dims[x_dims.size() - 1];
dim_out_vec[x_dims.size() - 1] = y_dims[y_dims.size() - 1];
} else {
dim_out_vec[x_dims.size() - 2] = x_dims[x_dims.size() - 1];
dim_out_vec[x_dims.size() - 1] = y_dims[y_dims.size() - 2];
}
} else if (x_dims.size() == 2 && y_dims.size() == 2) {
// x: [M, K], y: [K, N], out: [M, N]
// x: [M, K], y: [K, N], out: [M, N]
......
lite/operators/op_params.h
浏览文件 @ b35e89d6
......@@ -695,7 +695,7 @@ struct SliceParam {
std::vector<int> decrease_axis{};
};
/// ----------------------- shape operators ----------------------
/// ----------------------- squeeze operators ----------------------
struct SqueezeParam {
const lite::Tensor* X{};
lite::Tensor* Out{};
......@@ -719,7 +719,6 @@ struct MatMulParam {
bool transpose_Y{false};
float alpha{1.0f};
};
} // namespace operators
} // namespace lite
} // namespace paddle
lite/tests/kernels/matmul_compute_test.cc
浏览文件 @ b35e89d6
......@@ -152,29 +152,36 @@ class MatMulComputeTester : public arena::TestCase {
auto* out = scope->NewTensor(out_);
CHECK(out);
// todo alpha
std::vector<int64_t> dim_out_vec;
if (x_dims_.size() > 2 && y_dims_.size() >= 2) {
// x: [B, ..., M, K], y: [B, ..., K, N], out: [B, ..., M, N]
// x: [B, M, K], y: [K, N], out: [B, M, N]
if (x_transpose_ || y_transpose_) {
LOG(FATAL) << "not supported transpose for x and y.";
}
CHECK_EQ(x_dims_[x_dims_.size() - 1], y_dims_[y_dims_.size() - 2])
<< "not supported x_dims(" << x_dims_ << ") and y_dims(" << y_dims_
<< ")";
dim_out_vec.resize(x_dims_.size());
for (size_t i = 0; i < x_dims_.size() - 1; ++i) {
for (size_t i = 0; i < x_dims_.size() - 2; ++i) {
dim_out_vec[i] = x_dims_[i];
}
if (!x_transpose_ && !y_transpose_) {
dim_out_vec[x_dims_.size() - 2] = x_dims_[x_dims_.size() - 2];
dim_out_vec[x_dims_.size() - 1] = y_dims_[y_dims_.size() - 1];
} else if (!x_transpose_ && y_transpose_) {
dim_out_vec[x_dims_.size() - 2] = x_dims_[x_dims_.size() - 2];
dim_out_vec[x_dims_.size() - 1] = y_dims_[y_dims_.size() - 2];
} else if (x_transpose_ && !y_transpose_) {
dim_out_vec[x_dims_.size() - 2] = x_dims_[x_dims_.size() - 1];
dim_out_vec[x_dims_.size() - 1] = y_dims_[y_dims_.size() - 1];
} else {
dim_out_vec[x_dims_.size() - 2] = x_dims_[x_dims_.size() - 1];
dim_out_vec[x_dims_.size() - 1] = y_dims_[y_dims_.size() - 2];
}
out->Resize(dim_out_vec);
auto* out_data = out->mutable_data<float>();
int x_inner = x_dims_[x_dims_.size() - 2] * x_dims_[x_dims_.size() - 1];
if (y_dims_.size() > 2) {
int y_inner = y_dims_[y_dims_.size() - 2] * y_dims_[y_dims_.size() - 1];
int o_inner = x_dims_[x_dims_.size() - 2] * y_dims_[y_dims_.size() - 1];
int o_inner =
dim_out_vec[x_dims_.size() - 2] * dim_out_vec[x_dims_.size() - 1];
for (size_t i = 0; i < x_dims_.count(0, x_dims_.size() - 2); ++i) {
mul_low_efficiency(
DDim({x_dims_[x_dims_.size() - 2], x_dims_[x_dims_.size() - 1]}),
......@@ -187,7 +194,8 @@ class MatMulComputeTester : public arena::TestCase {
out_data + i * o_inner);
}
} else {
int o_inner = x_dims_[x_dims_.size() - 2] * y_dims_[1];
int o_inner =
dim_out_vec[x_dims_.size() - 2] * dim_out_vec[x_dims_.size() - 1];
for (size_t i = 0; i < x_dims_.count(0, x_dims_.size() - 2); ++i) {
mul_low_efficiency(
DDim({x_dims_[x_dims_.size() - 2], x_dims_[x_dims_.size() - 1]}),
......@@ -240,7 +248,7 @@ class MatMulComputeTester : public arena::TestCase {
out_data[i] += x_data[i * y_dims_[0] + j] * y_data[j] * alpha_;
}
}
} else if (x_dims_.size() == 1 && y_dims_.size() == 1) { // todo
} else if (x_dims_.size() == 1 && y_dims_.size() == 1) {
// x: [K], y: [K], out: [1]
if (x_dims_[0] == y_dims_[0] && x_transpose_ == false &&
y_transpose_ == false) {
......@@ -325,14 +333,40 @@ void test_matmul2x2_no_transform(Place place) {
}
}
void test_matmul2x2_transform(Place place) {
DDim x_dim({3, 2});
DDim y_dim({3, 2});
float alpha = 1.f;
std::unique_ptr<arena::TestCase> tester(
new MatMulComputeTester(place, "def", false, true, alpha, x_dim, y_dim));
void test_matmul2x2_x_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4}), DDim({2, 5})});
std::vector<DDim> y_dims({DDim({3, 2}), DDim({2, 1})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, false, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 2e-5);
arena.TestPrecision();
}
}
void test_matmul2x2_y_transpose(Place place) {
std::vector<DDim> x_dims({DDim({5, 2}), DDim({2, 5})});
std::vector<DDim> y_dims({DDim({3, 2}), DDim({1, 5})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", false, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 2e-5);
arena.TestPrecision();
}
}
void test_matmul2x2_transpose(Place place) {
std::vector<DDim> x_dims({DDim({6, 2}), DDim({5, 3})});
std::vector<DDim> y_dims({DDim({3, 6}), DDim({1, 5})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 5e-5);
arena.TestPrecision();
}
}
void test_matmul1x1_no_transpose(Place place) {
......@@ -366,9 +400,9 @@ void test_matmul_nx1(Place place) {
}
void test_matmul_nx2_1(Place place) {
DDim x_dim({3, 4, 2, 5});
DDim y_dim({5, 1});
float alpha = 1.5f;
DDim x_dim({1, 2, 2, 3});
DDim y_dim({3, 1});
float alpha = 1.f;
std::unique_ptr<arena::TestCase> tester(
new MatMulComputeTester(place, "def", false, false, alpha, x_dim, y_dim));
arena::Arena arena(std::move(tester), place, 2e-5);
......@@ -376,8 +410,8 @@ void test_matmul_nx2_1(Place place) {
}
void test_matmul_nx2_2(Place place) {
DDim x_dim({3, 4, 2, 5});
DDim y_dim({5, 3});
DDim x_dim({1, 2, 2, 3});
DDim y_dim({3, 3});
float alpha = 1.5f;
std::unique_ptr<arena::TestCase> tester(
new MatMulComputeTester(place, "def", false, false, alpha, x_dim, y_dim));
......@@ -385,14 +419,86 @@ void test_matmul_nx2_2(Place place) {
arena.TestPrecision();
}
void test_matmulnx2_x_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 5, 2})});
std::vector<DDim> y_dims({DDim({6, 2}), DDim({5, 1})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, false, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 2e-4);
arena.TestPrecision();
}
}
void test_matmulnx2_y_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 5, 2})});
std::vector<DDim> y_dims({DDim({6, 2}), DDim({1, 2})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", false, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 5e-5);
arena.TestPrecision();
}
}
void test_matmulnx2_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 4, 3}), DDim({5, 3, 3, 2})});
std::vector<DDim> y_dims({DDim({2, 4}), DDim({1, 3})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 5e-5);
arena.TestPrecision();
}
}
void test_matmul_nxn(Place place) {
DDim x_dim({3, 4, 2, 5});
DDim y_dim({3, 4, 5, 2});
float alpha = 1.5f;
std::unique_ptr<arena::TestCase> tester(
new MatMulComputeTester(place, "def", false, false, alpha, x_dim, y_dim));
arena::Arena arena(std::move(tester), place, 2e-5);
arena::Arena arena(std::move(tester), place, 1e-3);
arena.TestPrecision();
}
void test_matmulnxn_x_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 5, 2})});
std::vector<DDim> y_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 5, 1})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, false, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 1e-3);
arena.TestPrecision();
}
}
void test_matmulnxn_y_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 5, 2})});
std::vector<DDim> y_dims({DDim({3, 4, 6, 2}), DDim({5, 3, 1, 2})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", false, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 1e-3);
arena.TestPrecision();
}
}
void test_matmulnxn_transpose(Place place) {
std::vector<DDim> x_dims({DDim({3, 4, 4, 3}), DDim({5, 3, 3, 2})});
std::vector<DDim> y_dims({DDim({3, 4, 2, 4}), DDim({5, 3, 1, 3})});
std::vector<float> alphas({1.f, 2.f});
for (int i = 0; i < x_dims.size(); ++i) {
std::unique_ptr<arena::TestCase> tester(new MatMulComputeTester(
place, "def", true, true, alphas[i], x_dims[i], y_dims[i]));
arena::Arena arena(std::move(tester), place, 1e-3);
arena.TestPrecision();
}
}
TEST(Matmul2x2, precision) {
......@@ -401,11 +507,39 @@ TEST(Matmul2x2, precision) {
#endif
#ifdef LITE_WITH_ARM
Place place(TARGET(kARM));
// test_matmul2x2_transform(place);
test_matmul2x2_no_transform(place);
#endif
}
TEST(Matmul2x2_x_transpose, precision) {
#ifdef LITE_WITH_X86
Place place(TARGET(kX86));
#endif
#ifdef LITE_WITH_ARM
Place place(TARGET(kARM));
test_matmul2x2_x_transpose(place);
#endif
}
TEST(Matmul2x2_y_transpose, precision) {
#ifdef LITE_WITH_X86
Place place(TARGET(kX86));
#endif
#ifdef LITE_WITH_ARM
Place place(TARGET(kARM));
test_matmul2x2_y_transpose(place);
#endif
}
TEST(Matmul2x2_transpose, precision) {
#ifdef LITE_WITH_X86
Place place(TARGET(kX86));
#endif
#ifdef LITE_WITH_ARM
Place place(TARGET(kARM));
test_matmul2x2_transpose(place);
#endif
}
TEST(Matmul1x1, precision) {
#ifdef LITE_WITH_X86
Place place(TARGET(kX86));
......@@ -435,6 +569,9 @@ TEST(Matmulnx2, precision) {
Place place(TARGET(kARM));
test_matmul_nx2_1(place);
test_matmul_nx2_2(place);
test_matmulnx2_x_transpose(place);
test_matmulnx2_y_transpose(place);
test_matmulnx2_transpose(place);
#endif
}
......@@ -445,6 +582,9 @@ TEST(Matmulnxn, precision) {
#ifdef LITE_WITH_ARM
Place place(TARGET(kARM));
test_matmul_nxn(place);
test_matmulnxn_x_transpose(place);
test_matmulnxn_y_transpose(place);
test_matmulnxn_transpose(place);
#endif
}
......
Markdown is supported
0% 或 .
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 或