def verify(target="llvm", algorithm=nnpack.ConvolutionAlgorithm.AUTO, with_bias=True): if not tvm.module.enabled(target): pytest.skip("%s is not enabled..." % target) if not tvm.get_global_func("tvm.contrib.nnpack.fully_connected_inference", True): pytest.skip("extern function is not available") if not nnpack.is_available(): pytest.skip("nnpack is not available") ctx = tvm.cpu(0) transformed_kernel = nnpack.convolution_inference_weight_transform( kernel, algorithm=algorithm) output = nnpack.convolution_inference_without_weight_transform( data, transformed_kernel, bias if with_bias else None, [PAD, PAD, PAD, PAD], [STRIDE, STRIDE], algorithm=algorithm) s = tvm.create_schedule(output.op) f = tvm.build(s, [data, kernel, bias, output], target) na = np.random.uniform(size=dshape).astype(data.dtype) nb = np.random.uniform(size=kshape).astype(kernel.dtype) nc = np.random.uniform(size=bshape).astype(bias.dtype) if with_bias else np.zeros(bshape, dtype=bias.dtype) ta = tvm.nd.array(na, ctx) tb = tvm.nd.array(nb, ctx) tc = tvm.nd.array(nc, ctx) td = tvm.nd.array(np.zeros(oshape, dtype=output.dtype), ctx) f(ta, tb, tc, td) nd = np_conv(np.reshape(na, (BATCH, IC, IH, IW)), nb, PAD, STRIDE) + nc.reshape(1, bshape[0], 1, 1) tvm.testing.assert_allclose( td.asnumpy(), nd.reshape(BATCH, IC, IH, IW), rtol=1e-5)
def verify(target="llvm", algorithm=nnpack.ConvolutionAlgorithm.AUTO, with_bias=True): if not tvm.module.enabled(target): print("skip because %s is not enabled..." % target) return if not tvm.get_global_func("tvm.contrib.nnpack.fully_connected_inference", True): print("skip because extern function is not available") return if not nnpack.is_available(): return ctx = tvm.cpu(0) transformed_kernel = nnpack.convolution_inference_weight_transform( kernel, algorithm=algorithm) output = nnpack.convolution_inference_without_weight_transform( data, transformed_kernel, bias if with_bias else None, [PAD, PAD, PAD, PAD], [STRIDE, STRIDE], algorithm=algorithm) s = tvm.create_schedule(output.op) f = tvm.build(s, [data, kernel, bias, output], target) na = np.random.uniform(size=dshape).astype(data.dtype) nb = np.random.uniform(size=kshape).astype(kernel.dtype) nc = np.random.uniform(size=bshape).astype(bias.dtype) if with_bias else np.zeros(bshape, dtype=bias.dtype) ta = tvm.nd.array(na, ctx) tb = tvm.nd.array(nb, ctx) tc = tvm.nd.array(nc, ctx) td = tvm.nd.array(np.zeros(oshape, dtype=output.dtype), ctx) f(ta, tb, tc, td) nd = np_conv(np.reshape(na, (BATCH, IC, IH, IW)), nb, PAD, STRIDE) + nc.reshape(1, bshape[0], 1, 1) tvm.testing.assert_allclose( td.asnumpy(), nd.reshape(BATCH, IC, IH, IW), rtol=1e-5)
def conv2d_winograd_nnpack_weight_transform(kernel, convolution_algorithm, out_dtype): """Weight transformation for winograd Parameters ---------- kernel: Tensor The raw kernel tensor with layout "NCHW". Only 3x3 kernel is supported for now. convolution_algorithm: int The convolution algorithm for Winograd NNPACK. Returns ------- output : tvm.Tensor 4-D with shape [alpha, alpha, CO, CI] """ from tvm.contrib import nnpack return nnpack.convolution_inference_weight_transform( kernel, algorithm=convolution_algorithm, dtype=out_dtype)