From b67584a6e9a5083a4e6c5c87b9ac4b68d48a3647 Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Thu, 24 Jan 2019 17:14:58 +0000 Subject: [PATCH] jit benchmark use tensor test=develop --- paddle/fluid/operators/jit/CMakeLists.txt | 2 +- paddle/fluid/operators/jit/benchmark.cc | 107 ++++++++++++++-------- 2 files changed, 71 insertions(+), 38 deletions(-) diff --git a/paddle/fluid/operators/jit/CMakeLists.txt b/paddle/fluid/operators/jit/CMakeLists.txt index 262094f92..35775d7ec 100644 --- a/paddle/fluid/operators/jit/CMakeLists.txt +++ b/paddle/fluid/operators/jit/CMakeLists.txt @@ -21,5 +21,5 @@ endif() cc_library(jit_kernel_helper SRCS ${jit_kernel_cc_srcs} DEPS ${JIT_KERNEL_DEPS}) cc_test(jit_kernel_test SRCS test.cc DEPS jit_kernel_helper) if(NOT WIN32) - cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper device_tracer) + cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper device_tracer tensor) endif() diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc index 74d6a8724..186c37c56 100644 --- a/paddle/fluid/operators/jit/benchmark.cc +++ b/paddle/fluid/operators/jit/benchmark.cc @@ -18,6 +18,7 @@ #include #include "gflags/gflags.h" #include "glog/logging.h" +#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/operators/jit/kernels.h" #include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/platform/place.h" @@ -155,14 +156,22 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { LOG(INFO) << loginfos.str(); } +using Tensor = paddle::framework::Tensor; + template void BenchXYZNKernel() { for (int d : TestSizes()) { - std::vector x(d), y(d), z(d); - RandomVec(d, x.data()); - RandomVec(d, y.data()); - BenchAllImpls, PlaceType>(d, x.data(), y.data(), - z.data(), d); + Tensor x, y, z; + x.Resize({d}); + y.Resize({d}); + z.Resize({d}); + T* x_data = x.mutable_data(PlaceType()); + T* y_data = y.mutable_data(PlaceType()); + T* z_data = z.mutable_data(PlaceType()); + RandomVec(d, x_data); + RandomVec(d, y_data); + BenchAllImpls, PlaceType>(d, x.data(), + y.data(), z_data, d); } } @@ -170,9 +179,13 @@ template void BenchAXYNKernel() { for (int d : TestSizes()) { const T a = static_cast(3); - std::vector x(d), y(d); - RandomVec(d, x.data()); - BenchAllImpls, PlaceType>(d, &a, x.data(), y.data(), + Tensor x, y; + x.Resize({d}); + y.Resize({d}); + T* x_data = x.mutable_data(PlaceType()); + T* y_data = y.mutable_data(PlaceType()); + RandomVec(d, x_data); + BenchAllImpls, PlaceType>(d, &a, x.data(), y_data, d); } } @@ -180,9 +193,13 @@ void BenchAXYNKernel() { template void BenchXYNKernel() { for (int d : TestSizes()) { - std::vector x(d), y(d); - RandomVec(d, x.data()); - BenchAllImpls, PlaceType>(d, x.data(), y.data(), d); + Tensor x, y; + x.Resize({d}); + y.Resize({d}); + T* x_data = x.mutable_data(PlaceType()); + T* y_data = y.mutable_data(PlaceType()); + RandomVec(d, x_data); + BenchAllImpls, PlaceType>(d, x.data(), y_data, d); } } @@ -192,16 +209,23 @@ void BenchLSTMKernel() { for (int d : TestSizes()) { const jit::lstm_attr_t attr(d, jit::kVSigmoid, jit::kVTanh, jit::kVTanh, use_peephole); - std::vector x(4 * d), ct_1(d), ct(d), ht(d), wp(3 * d), checked(2 * d); - RandomVec(4 * d, x.data(), -2.f, 2.f); - RandomVec(3 * d, wp.data(), -2.f, 2.f); - RandomVec(d, ct_1.data(), -2.f, 2.f); - const T* ct_1_data = ct_1.data(); - const T* wp_data = wp.data(); - T* x_data = x.data(); - T* checked_data = checked.data(); - T* ct_data = ct.data(); - T* ht_data = ht.data(); + Tensor x, ct_1, ct, ht, wp, checked; + x.Resize({4 * d}); + ct_1.Resize({d}); + ct.Resize({d}); + ht.Resize({d}); + wp.Resize({3 * d}); + checked.Resize({2 * d}); + auto place = PlaceType(); + RandomVec(x.numel(), x.mutable_data(place), -2.f, 2.f); + RandomVec(wp.numel(), wp.mutable_data(place), -2.f, 2.f); + RandomVec(ct_1.numel(), ct_1.mutable_data(place), -2.f, 2.f); + const T* ct_1_data = ct_1.data(); + const T* wp_data = wp.data(); + T* x_data = x.mutable_data(place); + T* checked_data = checked.mutable_data(place); + T* ct_data = ct.mutable_data(place); + T* ht_data = ht.mutable_data(place); jit::lstm_t step; step.gates = x_data; step.ct_1 = ct_1_data; @@ -220,12 +244,16 @@ template void BenchGRUKernel() { for (int d : TestSizes()) { const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh); - std::vector x(3 * d), ht_1(d), ht(d); - RandomVec(3 * d, x.data(), -2.f, 2.f); - RandomVec(d, ht_1.data(), -2.f, 2.f); - const T* ht_1_data = ht_1.data(); - T* x_data = x.data(); - T* ht_data = ht.data(); + auto place = PlaceType(); + Tensor x, ht_1, ht; + x.Resize({3 * d}); + ht_1.Resize({d}); + ht.Resize({d}); + RandomVec(3 * d, x.mutable_data(place), -2.f, 2.f); + RandomVec(d, ht_1.mutable_data(place), -2.f, 2.f); + const T* ht_1_data = ht_1.data(); + T* x_data = x.mutable_data(place); + T* ht_data = ht.mutable_data(place); jit::gru_t step; step.gates = x_data; step.ht_1 = ht_1_data; @@ -243,10 +271,12 @@ void BenchSeqPoolKernel() { jit::seq_pool_attr_t attr(w, type); for (int h : TestSizes()) { attr.h = h; - std::vector x(h * w), y(w); - RandomVec(h * w, x.data(), -2.f, 2.f); - const T* x_data = x.data(); - T* y_data = y.data(); + Tensor x, y; + x.Resize({h * w}); + y.Resize({w}); + RandomVec(h * w, x.mutable_data(PlaceType()), -2.f, 2.f); + const T* x_data = x.data(); + T* y_data = y.mutable_data(PlaceType()); BenchAllImpls, PlaceType>(attr, x_data, y_data, &attr); } @@ -259,12 +289,15 @@ void BenchMatMulKernel() { for (int m : {1, 2, 3, 4}) { for (int n : TestSizes()) { for (int k : TestSizes()) { - std::vector a(m * k), b(k * n), c(m * n); - RandomVec(m * k, a.data(), -2.f, 2.f); - RandomVec(k * n, b.data(), -2.f, 2.f); - const T* a_data = a.data(); - const T* b_data = b.data(); - T* c_data = c.data(); + Tensor a, b, c; + a.Resize({m * k}); + b.Resize({k * n}); + c.Resize({m * n}); + RandomVec(m * k, a.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(k * n, b.mutable_data(PlaceType()), -2.f, 2.f); + const T* a_data = a.data(); + const T* b_data = b.data(); + T* c_data = c.mutable_data(PlaceType()); BenchAllImpls, PlaceType>(k, a_data, b_data, c_data, m, n, k); } -- GitLab