new backend gemv speedhack using vector-vector dot

theano/gpuarray/blas.py

@@ -9,6 +9,7 @@ from theano.compile import optdb
 from theano.gof import LocalOptGroup
 from theano.tensor.basic import as_tensor_variable
 from theano.tensor.opt import in2out
+from theano.tensor.var import TensorConstant
 
 from .basic_ops import (GpuArrayType, CGpuKernelBase,
                         as_gpuarray_variable, gpu_contiguous, infer_context_name)
@@ -52,6 +53,15 @@ class GpuGemv(BlasOp):
         y = as_gpuarray_variable(y, ctx_name)
         alpha = as_tensor_variable(alpha).astype('float64')
         beta = as_tensor_variable(beta).astype('float64')
+
+        # if alpha==1. and beta==0., we add runtime check
+        # for possible speed up using vector-vector dot as
+        # gemv tend to be slower for vector-vector dot
+        self._use_rt_dot_check = False
+        if all(map(lambda v: isinstance(v, TensorConstant), [alpha, beta])):
+            if alpha.value == 1. and beta.value == 0.:
+                self._use_rt_dot_check = True
+
         assert alpha.ndim == 0
         assert beta.ndim == 0
         assert A.ndim == 2
@@ -65,6 +75,7 @@ class GpuGemv(BlasOp):
         inplace = self.inplace
         if inplace and y.strides[0] < 0:
             inplace = False
+        print(alpha, beta)
         out_storage[0][0] = blas.gemv(alpha, A, x, beta, y,
                                       overwrite_y=inplace)
 
@@ -99,7 +110,33 @@ class GpuGemv(BlasOp):
             PyErr_SetString(PyExc_RuntimeError, "Memset failed");
             %(fail)s
           }
-        } else if (pygpu_blas_rgemv(cb_no_trans,
+        } """ % vars
+        if self._use_rt_dot_check:
+            # temporary hack A to 1D for vector-vector dot
+            code += """
+            else if (%(A)s->ga.dimensions[0]==1) {
+                %(out)s->ga.nd = 0;
+                %(A)s->ga.nd = 1;
+                if (%(A)s->ga.flags & GA_C_CONTIGUOUS) {
+                    %(A)s->ga.strides[0] ^= %(A)s->ga.strides[1];
+                    %(A)s->ga.strides[1] ^= %(A)s->ga.strides[0];
+                    %(A)s->ga.strides[0] ^= %(A)s->ga.strides[1];
+                }
+                %(A)s->ga.dimensions[0] = %(A)s->ga.dimensions[1];
+                if (pygpu_blas_rdot(%(x)s, %(A)s, %(y)s, 0) == -1) {
+                    %(fail)s
+                }
+                %(A)s->ga.dimensions[0] = 1;
+                if (%(A)s->ga.flags & GA_C_CONTIGUOUS) {
+                    %(A)s->ga.strides[0] ^= %(A)s->ga.strides[1];
+                    %(A)s->ga.strides[1] ^= %(A)s->ga.strides[0];
+                    %(A)s->ga.strides[0] ^= %(A)s->ga.strides[1];
+                }
+                %(A)s->ga.nd = 2;
+                %(out)s->ga.nd = 1;
+            } """ % vars
+        code += """
+        else if (pygpu_blas_rgemv(cb_no_trans,
                              ((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
                              %(A)s, %(x)s,
                              ((dtype_%(beta)s *)PyArray_DATA(%(beta)s))[0],