Remove the special code that would move float16 dot to the gpu since

theano/gpuarray/opt.py

@@ -1134,27 +1134,6 @@ def local_gpua_gemmbatch(op, context_name, inputs, outputs):
     return gpugemmbatch_no_inplace(c, 1.0, a, b, 0.0)
 
 
-@register_opt('fast_compile')
-@op_lifter([tensor.basic.Dot])
-@register_opt2([tensor.basic.Dot], 'fast_compile')
-def local_gpua_hgemm(op, context_name, inputs, outputs):
-    from theano.sandbox.cuda import nvcc_compiler
-    if nvcc_compiler.nvcc_version < '7.5':
-        _logger.warning("Not performing dot of float16 on the GPU since "
-                        "cuda 7.5 is not available. Updating could speed up "
-                        "your code.")
-        return
-    A = inputs[0]
-    B = inputs[1]
-    if (A.ndim == 2 and B.ndim == 2 and
-            A.dtype == 'float16' and B.dtype == 'float16'):
-        fgraph = outputs[0].fgraph
-        C = gpu_alloc_empty(context_name, dtype='float16')(
-            shape_i(A, 0, fgraph),
-            shape_i(B, 1, fgraph))
-        return gpugemm_no_inplace(C, 1.0, A, B, 0.0)
-
-
 @register_opt()
 @alpha_merge(GpuGemm, alpha_in=1, beta_in=4)
 def local_gpua_gemm_alpha_merge(node, *inputs):

theano/gpuarray/tests/test_blas.py

@@ -130,52 +130,3 @@ GpuDot22Tester = makeTester(
         # test9=[rand(0, 0), rand(0, 0)],
     )
 )
-
-
-def test_hgemm_swap():
-    from theano.sandbox.cuda import nvcc_compiler
-    if nvcc_compiler.nvcc_version < '7.5':
-        raise SkipTest("SgemmEx is only avaialble on cuda 7.5+")
-
-    v = tensor.vector(dtype='float16')
-    m = tensor.matrix(dtype='float16')
-    m2 = tensor.matrix(dtype='float16')
-    m32 = tensor.matrix(dtype='float32')
-
-    # test that we don't try to replace anything but matrix x matrix in float16
-    f = theano.function([v, m], tensor.dot(v, m), mode=mode_with_gpu)
-    assert len([node for node in f.maker.fgraph.apply_nodes
-                if isinstance(node.op, GpuGemm)]) == 0
-
-    f = theano.function([m32, m], tensor.dot(m32, m), mode=mode_with_gpu)
-    assert len([node for node in f.maker.fgraph.apply_nodes
-                if isinstance(node.op, GpuGemm)]) == 0
-
-    f = theano.function([m, m2], tensor.dot(m, m2), mode=mode_with_gpu)
-    assert len([node for node in f.maker.fgraph.apply_nodes
-                if isinstance(node.op, GpuGemm)]) == 1
-
-    v1 = numpy.random.random((3, 4)).astype('float16')
-    v2 = numpy.random.random((4, 2)).astype('float16')
-
-    of = f(v1, v2)
-    on = numpy.dot(v1, v2)
-
-    utt.assert_allclose(of, on)
-
-
-def test_hgemm_alpha_output_merge():
-    from theano.sandbox.cuda import nvcc_compiler
-    if nvcc_compiler.nvcc_version < '7.5':
-        raise SkipTest("SgemmEx is only avaialble on cuda 7.5+")
-
-    m1 = tensor.matrix(dtype='float16')
-    m2 = tensor.matrix(dtype='float16')
-
-    b = tensor.matrix(dtype='float16')
-
-    hgemm = numpy.asarray(0.05, dtype='float16') * (tensor.dot(m1, m2) + b)
-
-    f = theano.function([m1, m2, b], hgemm, mode=mode_with_gpu)
-    # there should be 3 gpu_from_host, 1 hgemm and 1 host_from_gpu
-    assert len(f.maker.fgraph.apply_nodes) == 5