fix doc and jit. test=develop

paddle/fluid/API.spec

@@ -86,7 +86,7 @@ paddle.fluid.layers.conv2d (ArgSpec(args=['input', 'num_filters', 'filter_size',
 paddle.fluid.layers.conv3d (ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None)), ('document', '37042620f9bd3a2da6e5d3138b2f724b'))
 paddle.fluid.layers.sequence_pool (ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,)), ('document', 'a194fb80614023f543df3949fbd0d0b8'))
 paddle.fluid.layers.sequence_softmax (ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '19ef6f9cdd27feac8a1ae060f19c10b4'))
-paddle.fluid.layers.softmax (ArgSpec(args=['input', 'use_cudnn', 'name', 'axis'], varargs=None, keywords=None, defaults=(False, None, -1)), ('document', '502bad9e8bc7ef24817d0d4b20f61df3'))
+paddle.fluid.layers.softmax (ArgSpec(args=['input', 'use_cudnn', 'name', 'axis'], varargs=None, keywords=None, defaults=(False, None, -1)), ('document', '59b1c6bf2f0fa9dc649c85fef3a3b2ea'))
 paddle.fluid.layers.pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)), ('document', 'bbd84e855e660cd1084bb71a2fd0cdaa'))
 paddle.fluid.layers.pool3d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)), ('document', '043de7333b79ee0ac55053c14ed81625'))
 paddle.fluid.layers.adaptive_pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)), ('document', '859b887174d06f361658f69cb7c06d95'))

paddle/fluid/operators/jit/kernel_base.h

@@ -38,6 +38,8 @@ typedef enum {
   kNCHW16CMulNC,
   kSeqPool,
   kSoftmax,
+  kStrideASum,
+  kStrideScal,
   kVAdd,
   kVAddBias,
   kVAddRelu,
@@ -53,8 +55,6 @@ typedef enum {
   kVSquare,
   kVSub,
   kVTanh,
-  kStrideASum,
-  kStrideScal,
 } KernelType;
 
 typedef enum {

paddle/fluid/operators/jit/more/mix/mix.cc

@@ -50,11 +50,12 @@ void VTanh(const T* x, T* y, int n) {
   compute_addbias(&b, y, y, n);
 }
 
+// remain is the product of dimension shapes after the axis dimension
 void Softmax(const T* x, T* y, int n, int bs, int remain) {
   auto compute_hmax = KernelFuncs<HMaxTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_hsum = KernelFuncs<HSumTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_vscal = KernelFuncs<VScalTuple<T>, CPUPlace>::Cache().At(n);
-  auto compute_stridesum =
+  auto compute_strideasum =
       KernelFuncs<StrideASumTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_stridescal =
       KernelFuncs<StrideScalTuple<T>, CPUPlace>::Cache().At(n);
@@ -74,7 +75,7 @@ void Softmax(const T* x, T* y, int n, int bs, int remain) {
       compute_vscal(&scalar, y, y, n);
     } else {
       for (int j = 0; j < remain; ++j) {
-        compute_stridesum(&y[j], &scalar, n, remain);
+        compute_strideasum(&y[j], &scalar, n, remain);
         scalar = static_cast<T>(1) / scalar;
         compute_stridescal(&scalar, &y[j], &y[j], n, remain);
       }

paddle/fluid/operators/jit/more/mkl/mkl.h

@@ -134,6 +134,7 @@ void StrideASum(const T* x, T* res, int n, int stride);
 template <typename T>
 void StrideScal(const T* a, const T* x, T* y, int n, int stride);
 
+// remain is the product of dimension shapes after the axis dimension
 template <typename T>
 void Softmax(const T* x, T* y, int n, int bs, int remain = 1) {
   std::vector<T> entities(bs);

paddle/fluid/operators/jit/refer/refer.h

@@ -432,6 +432,7 @@ void StrideScal(const T* a, const T* x, T* y, int n, int stride) {
 
 // y = e^(x - max(x))
 // y = y / sum(y)
+// remain is the product of dimension shapes after the axis dimension
 template <typename T>
 void Softmax(const T* x, T* y, int n, int bs = 1, int remain = 1) {
   for (int i = 0; i < bs; ++i) {

paddle/fluid/operators/jit/test.cc

@@ -798,10 +798,8 @@ template <typename KernelTuple, typename PlaceType>
 void TestKernelStrideScal() {
   using T = typename KernelTuple::data_type;
   VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type);
-  // for (int d : TestSizes()) {
-  //   for (int m : {1, 2, 3}) { // stride
-  for (int d : {4}) {
-    for (int m : {2}) {  // stride
+  for (int d : TestSizes()) {
+    for (int m : {1, 2, 3}) { // stride
       if (m > d || d % m != 0) {
         continue;
       }

python/paddle/fluid/layers/nn.py

@@ -1826,7 +1826,7 @@ def softmax(input, use_cudnn=False, name=None, axis=-1):
 
     The dimension :attr:`axis` of the input tensor will be permuted to the last.
     Then the input tensor will be logically flattened to a 2-D matrix. The matrix's
-    second dimension(row length) is as same as the dimension :attr:`axis` of the input
+    second dimension(row length) is the same as the dimension :attr:`axis` of the input
     tensor, and the first dimension(column length) is the product of all other
     dimensions of the input tensor. For each row of the matrix, the softmax operator
     squashes the K-dimensional(K is the width of the matrix, which is also the size
@@ -1864,7 +1864,10 @@ def softmax(input, use_cudnn=False, name=None, axis=-1):
         .. code-block:: python
 
              fc = fluid.layers.fc(input=x, size=10)
+             # perform softmax in the second dimension
              softmax = fluid.layers.softmax(input=fc, axis=1)
+             # perform softmax in the last dimension
+             softmax = fluid.layers.softmax(input=fc, axis=-1)
 
     """
     helper = LayerHelper('softmax', **locals())