[CPU] refine cpu softmax bwd (#17534)

* refine softmax fwd

test=develop

* refine cpu softmax bwd

test=develop

* fix batch size

test=develop

* fix compile issue with gpu

test=develop

* add value clip

paddle/fluid/operators/math/cpu_vec.h

@@ -176,6 +176,79 @@ inline void vec_sum<float, platform::avx>(const size_t n, const float* x,
 #endif
 }
 
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
+inline void vec_mul(const size_t n, const T* x, const T* y, T* z) {
+  for (size_t i = 0; i < n; ++i) {
+    z[i] = x[i] * y[i];
+  }
+}
+
+template <>
+inline void vec_mul<float, platform::avx>(const size_t n, const float* x,
+                                          const float* y, float* z) {
+#ifdef __AVX__
+  constexpr unsigned int block = YMM_FLOAT_BLOCK;
+  if (n < block) {
+    vec_mul<float, platform::isa_any>(n, x, y, z);
+    return;
+  }
+
+  unsigned int i = 0, end = 0;
+  end = n & ~(block - 1);
+  for (i = 0; i < end; i += block) {
+    _mm256_storeu_ps(
+        z + i, _mm256_mul_ps(_mm256_loadu_ps(x + i), _mm256_loadu_ps(y + i)));
+  }
+
+  for (; i < n; i++) {
+    z[i] = x[i] * y[i];
+  }
+#else
+  vec_mul<float, platform::isa_any>(n, x, y, z);
+#endif
+}
+
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
+inline void vec_mul_reduce(const size_t n, const T* x, const T* y, T* z) {
+  z[0] = x[0] * y[0];
+  for (size_t i = 1; i < n; ++i) {
+    z[0] += x[i] * y[i];
+  }
+}
+
+template <>
+inline void vec_mul_reduce<float, platform::avx>(const size_t n, const float* x,
+                                                 const float* y, float* z) {
+#ifdef __AVX__
+  constexpr unsigned int block = YMM_FLOAT_BLOCK;
+  if (n < block) {
+    vec_mul_reduce<float, platform::isa_any>(n, x, y, z);
+    return;
+  }
+
+  unsigned int i = 0, end = 0;
+  z[0] = 0.f;
+
+  end = n & ~(block - 1);
+  __m256 tmp = _mm256_setzero_ps();
+  for (i = 0; i < end; i += block) {
+    tmp = _mm256_add_ps(
+        tmp, _mm256_mul_ps(_mm256_loadu_ps(x + i), _mm256_loadu_ps(y + i)));
+  }
+
+  __m256 hsum = _mm256_hadd_ps(tmp, tmp);
+  hsum = _mm256_add_ps(hsum, _mm256_permute2f128_ps(hsum, hsum, 0x1));
+  _mm_store_ss(z, _mm_hadd_ps(_mm256_castps256_ps128(hsum),
+                              _mm256_castps256_ps128(hsum)));
+
+  for (; i < n; i++) {
+    z[0] += x[i] * y[i];
+  }
+#else
+  vec_mul_reduce<float, platform::isa_any>(n, x, y, z);
+#endif
+}
+
 template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_bias_sub(const int n, const T a, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {

paddle/fluid/operators/math/cpu_vec_test.cc

@@ -199,6 +199,70 @@ TEST(CpuVecTest, vec_clip) {
                        vec_clip<double, platform::isa_any>);
 }
 
+template <typename T>
+void compare_mul(
+    size_t n, std::function<void(const size_t, const T*, const T*, T*)> tgt,
+    std::function<void(const size_t, const T*, const T*, T*)> ref) {
+  std::vector<T> x(n), y(n);
+  std::vector<T> ztgt(n), zref(n);
+
+  RandomVec<T>(n, x.data(), static_cast<T>(-2), static_cast<T>(2));
+  RandomVec<T>(n, y.data(), static_cast<T>(-2), static_cast<T>(2));
+
+  const T* x_data = x.data();
+  const T* y_data = y.data();
+  T* ztgt_data = ztgt.data();
+  T* zref_data = zref.data();
+
+  tgt(n, x_data, y_data, ztgt_data);
+  ref(n, x_data, y_data, zref_data);
+  for (size_t i = 0; i < n; ++i) {
+    EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+  }
+}
+
+TEST(CpuVecTest, vec_mul) {
+  namespace platform = paddle::platform;
+  using namespace paddle::operators::math;  // NOLINT
+  for (size_t sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    compare_mul<float>(sz, vec_mul<float>, vec_mul<float, platform::isa_any>);
+    compare_mul<float>(sz, vec_mul<float, platform::avx>,
+                       vec_mul<float, platform::isa_any>);
+  }
+  compare_mul<double>(30U, vec_mul<double>, vec_mul<double, platform::isa_any>);
+}
+
+template <typename T>
+void compare_mul_reduce(
+    size_t n, std::function<void(const size_t, const T*, const T*, T*)> tgt,
+    std::function<void(const size_t, const T*, const T*, T*)> ref) {
+  std::vector<T> x(n), y(n);
+  T ztgt_data, zref_data;
+
+  RandomVec<T>(n, x.data(), static_cast<T>(-2), static_cast<T>(2));
+  RandomVec<T>(n, y.data(), static_cast<T>(-2), static_cast<T>(2));
+
+  const T* x_data = x.data();
+  const T* y_data = y.data();
+
+  tgt(n, x_data, y_data, &ztgt_data);
+  ref(n, x_data, y_data, &zref_data);
+  EXPECT_NEAR(ztgt_data, zref_data, 1e-3);
+}
+
+TEST(CpuVecTest, vec_mul_reduce) {
+  namespace platform = paddle::platform;
+  using namespace paddle::operators::math;  // NOLINT
+  for (size_t sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    compare_mul_reduce<float>(sz, vec_mul_reduce<float>,
+                              vec_mul_reduce<float, platform::isa_any>);
+    compare_mul_reduce<float>(sz, vec_mul_reduce<float, platform::avx>,
+                              vec_mul_reduce<float, platform::isa_any>);
+  }
+  compare_mul_reduce<double>(30U, vec_mul_reduce<double>,
+                             vec_mul_reduce<double, platform::isa_any>);
+}
+
 template <typename T>
 void TestInplace(const int n, std::function<void(const int, const T*, T*)> tgt,
                  std::function<void(const int, const T*, T*)> ref) {

paddle/fluid/operators/math/softmax.h

@@ -27,7 +27,7 @@ class SoftmaxFunctor {
                   const framework::Tensor* X, framework::Tensor* Y);
 };
 
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, typename Enable = void>
 class SoftmaxGradFunctor {
  public:
   void operator()(const DeviceContext& context, const int axis_dim,

paddle/fluid/operators/math/softmax_impl.h

@@ -140,16 +140,16 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
 };
 
 template <typename DeviceContext, typename T>
-void SoftmaxGradFunctor<DeviceContext, T>::operator()(
-    const DeviceContext& context, const int axis_dim,
-    const framework::Tensor* y, const framework::Tensor* y_grad,
-    framework::Tensor* x_grad) {
+void SoftmaxGradEigen(const DeviceContext& context, const int axis_dim,
+                      const framework::Tensor* y,
+                      const framework::Tensor* y_grad,
+                      framework::Tensor* x_grad) {
   auto softmax = EigenMatrix<T>::From(*y);
   auto softmax_grad = EigenMatrix<T>::From(*y_grad);
   auto logits_grad = EigenMatrix<T>::From(*x_grad);
 
-  const int kBatchDim = 0;
-  const int kClassDim = 1;
+  constexpr int kBatchDim = 0;
+  constexpr int kClassDim = 1;
 
   const int batch_size = softmax.dimension(kBatchDim);
   const int num_classes = softmax.dimension(kClassDim);
@@ -169,6 +169,48 @@ void SoftmaxGradFunctor<DeviceContext, T>::operator()(
   logits_grad.device(*context.eigen_device()) = (softmax_grad - dot) * softmax;
 }
 
+template <typename DeviceContext, typename T, typename Enable>
+void SoftmaxGradFunctor<DeviceContext, T, Enable>::operator()(
+    const DeviceContext& context, const int axis_dim,
+    const framework::Tensor* y, const framework::Tensor* y_grad,
+    framework::Tensor* x_grad) {
+  SoftmaxGradEigen<DeviceContext, T>(context, axis_dim, y, y_grad, x_grad);
+}
+
+template <typename DeviceContext, typename T>
+class SoftmaxGradFunctor<DeviceContext, T, enable_if_CPU<DeviceContext>> {
+ public:
+  void operator()(const DeviceContext& context, const int axis_dim,
+                  const framework::Tensor* y, const framework::Tensor* y_grad,
+                  framework::Tensor* x_grad) {
+    auto out_dims = y->dims();
+    constexpr int kBatchDim = 0;
+    constexpr int kClassDim = 1;
+    const int num_classes = out_dims[kClassDim];
+    const int batch_size = out_dims[kBatchDim];
+    const int num_remain = num_classes / axis_dim;
+
+    if (num_remain == 1 && platform::MayIUse(platform::avx)) {
+      const T* out_data = y->data<T>();
+      const T* out_grad = y_grad->data<T>();
+      T* in_grad = x_grad->data<T>();
+      for (int bs = 0; bs < batch_size; ++bs) {
+        T scalar;
+        vec_mul_reduce<T, platform::avx>(num_classes, out_grad, out_data,
+                                         &scalar);
+        scalar *= static_cast<T>(-1);
+        vec_add_bias<T, platform::avx>(num_classes, scalar, out_grad, in_grad);
+        vec_mul<T, platform::avx>(num_classes, out_data, in_grad, in_grad);
+        out_data += num_classes;
+        out_grad += num_classes;
+        in_grad += num_classes;
+      }
+    } else {
+      SoftmaxGradEigen<DeviceContext, T>(context, axis_dim, y, y_grad, x_grad);
+    }
+  }
+};
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle