From e5ce9659522553e373227d760a1b993dfe337e44 Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Fri, 19 Oct 2018 11:09:33 +0800
Subject: [PATCH] refine and add eltadd_relu unit test

---
 paddle/fluid/operators/math/fc_compute.h      |  2 +-
 .../fluid/operators/math/jit_kernel_blas.cc   |  3 -
 .../fluid/operators/math/jit_kernel_test.cc   | 57 +++++++++++++++++++
 3 files changed, 58 insertions(+), 4 deletions(-)

diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h
index 2d7e877a7..87220d401 100644
--- a/paddle/fluid/operators/math/fc_compute.h
+++ b/paddle/fluid/operators/math/fc_compute.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/jit_kernel.h"  // TODO(TJ): add deps
+#include "paddle/fluid/operators/math/jit_kernel.h"
 
 DECLARE_int32(paddle_num_threads);
 
diff --git a/paddle/fluid/operators/math/jit_kernel_blas.cc b/paddle/fluid/operators/math/jit_kernel_blas.cc
index a486a0ca8..c88b17b01 100644
--- a/paddle/fluid/operators/math/jit_kernel_blas.cc
+++ b/paddle/fluid/operators/math/jit_kernel_blas.cc
@@ -447,20 +447,17 @@ class VAddReluKernelImpl : public VAddReluKernel<T> {
 #ifdef __AVX__
 INTRI8_FLOAT(jit::avx);
 INTRI16_FLOAT(jit::avx);
-INTRI_COMMON_FLOAT(jit::avx, kGT8LT16);
 INTRI_COMMON_FLOAT(jit::avx, kGT16);
 #endif
 #ifdef __AVX2__
 INTRI8_FLOAT(jit::avx2);
 INTRI16_FLOAT(jit::avx2);
-INTRI_COMMON_FLOAT(jit::avx2, kGT8LT16);
 INTRI_COMMON_FLOAT(jit::avx2, kGT16);
 #endif
 #ifdef __AVX512F__
 // TODO(TJ): refine avx512
 INTRI8_FLOAT(jit::avx512f);
 INTRI16_FLOAT(jit::avx512f);
-INTRI_COMMON_FLOAT(jit::avx512f, kGT8LT16);
 INTRI_COMMON_FLOAT(jit::avx512f, kGT16);
 #endif
 
diff --git a/paddle/fluid/operators/math/jit_kernel_test.cc b/paddle/fluid/operators/math/jit_kernel_test.cc
index 7fdd1c6b7..c9e6ab740 100644
--- a/paddle/fluid/operators/math/jit_kernel_test.cc
+++ b/paddle/fluid/operators/math/jit_kernel_test.cc
@@ -712,6 +712,63 @@ TEST(JitKernel, vadd) {
   }
 }
 
+void vaddrelu_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] + y[i];
+    z[i] = z[i] > 0 ? z[i] : 0;
+  }
+}
+void vaddrelu_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VAddKernel<float>>& vadd,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VReluKernel<float>>& vrelu,
+    const float* x, const float* y, float* z) {
+  vadd->Compute(x, y, z);
+  vrelu->Compute(z, z);
+}
+
+TEST(JitKernel, vaddrelu) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddReluKernel<float>>(d);
+    const auto& vadd =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+    const auto& vrelu =
+        jit::KernelPool::Instance().template Get<jit::VReluKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vaddrelu_better(vadd, vrelu, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better takes: " << (tmkle - tmkls) / repeat << " us, "
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
 TEST(JitKernel, pool) {
   namespace jit = paddle::operators::math::jitkernel;
   const int frame_size = 4;
-- 
GitLab