comaniac commented on a change in pull request #9986:
URL: https://github.com/apache/tvm/pull/9986#discussion_r789902697
##########
File path: python/tvm/topi/cuda/conv2d.py
##########
@@ -123,3 +123,22 @@ def conv2d_cudnn(
def schedule_conv2d_cudnn(cfg, outs):
"""Create the schedule for conv2d_cudnn"""
return generic.schedule_extern(outs)
+
+
+def conv2d_backward_weight(
+ dy, x, kernel_size, padding, stride, dilation, groups, layout, output_dtype
+):
+ """Compute conv2d wgrad using CuDNN library"""
+ assert layout in ["NCHW", "NHWC"]
+ return cudnn.conv_backward_filter(
+ dy,
+ x,
+ kernel_size,
+ padding,
+ stride,
+ dilation,
+ conv_mode=1,
+ tensor_format=0 if layout == "NCHW" else 1,
+ conv_dtype=output_dtype,
+ groups=groups,
+ )
Review comment:
This might confuse people IMHO. Maybe name this function
`conv2d_backward_weight_cudnn` as other conv2ds?
##########
File path: tests/python/contrib/test_cudnn.py
##########
@@ -236,6 +236,148 @@ def test_softmax():
verify_softmax_4d((1, 16, 256, 256), "float64", log_softmax=True)
+def verify_conv2d_backward_data(data_dtype, conv_dtype, tensor_format=0,
tol=1e-5):
+ batch = 3
+ in_channel = 4
+ out_channel = 16
+ filter_h, filter_w = 3, 3
+ pad_h, pad_w = 1, 1
+ stride_h, stride_w = 1, 1
+ height, width = 32, 32
+
+ if tensor_format == 0:
+ xshape = [batch, in_channel, height, width]
+ wshape = [out_channel, in_channel, filter_h, filter_w]
+ oshape = xshape
+ oshape[1] = out_channel
+ ref_func = tvm.topi.testing.conv2d_transpose_nchw_python
+ else:
+ xshape = [batch, height, width, in_channel]
+ wshape = [out_channel, filter_h, filter_w, in_channel]
+ oshape = xshape
+ oshape[3] = out_channel
+ ref_func = lambda dy_np, w_np, strides, padding, out_pad:
tvm.topi.testing.conv2d_transpose_nhwc_python(
+ dy_np, np.transpose(w_np, [1, 2, 3, 0]), "HWOI", strides, padding,
out_pad
+ )
+
+ dy_np = np.random.uniform(-1, 1, oshape).astype(data_dtype)
+ w_np = np.random.uniform(-1, 1, wshape).astype(data_dtype)
+
+ if data_dtype == "float16":
+ dx_np = ref_func(
+ dy_np.astype("float32"),
+ w_np.astype("float32"),
+ (stride_h, stride_w),
+ (pad_h, pad_w),
+ (0, 0),
+ )
+ dx_np = dx_np.astype("float16")
+ else:
+ dx_np = ref_func(dy_np, w_np, (stride_h, stride_w), (pad_h, pad_w),
(0, 0))
+
+ dy = te.placeholder(oshape, name="dy", dtype=data_dtype)
+ w = te.placeholder(wshape, name="dw", dtype=data_dtype)
+ dx = cudnn.conv_backward_data(
+ dy,
+ w,
+ [pad_h, pad_w],
+ [stride_h, stride_w],
+ [1, 1],
+ conv_mode=1,
+ tensor_format=tensor_format,
+ conv_dtype=conv_dtype,
+ groups=1,
+ )
+
+ s = te.create_schedule(dx.op)
+
+ dev = tvm.cuda(0)
+ f = tvm.build(s, [dy, w, dx], "cuda --host=llvm",
name="conv2d_backward_data")
+
+ dy = tvm.nd.array(dy_np, dev)
+ w = tvm.nd.array(w_np, dev)
+ dx = tvm.nd.array(dx_np, dev)
+
+ f(dy, w, dx)
+ print(np.max(np.abs(dx.numpy() - dx_np)))
+ print(np.mean(np.abs(dx.numpy() - dx_np)))
Review comment:
Remove?
##########
File path: tests/python/contrib/test_cudnn.py
##########
@@ -236,6 +236,148 @@ def test_softmax():
verify_softmax_4d((1, 16, 256, 256), "float64", log_softmax=True)
+def verify_conv2d_backward_data(data_dtype, conv_dtype, tensor_format=0,
tol=1e-5):
+ batch = 3
+ in_channel = 4
+ out_channel = 16
+ filter_h, filter_w = 3, 3
+ pad_h, pad_w = 1, 1
+ stride_h, stride_w = 1, 1
+ height, width = 32, 32
+
+ if tensor_format == 0:
+ xshape = [batch, in_channel, height, width]
+ wshape = [out_channel, in_channel, filter_h, filter_w]
+ oshape = xshape
+ oshape[1] = out_channel
+ ref_func = tvm.topi.testing.conv2d_transpose_nchw_python
+ else:
+ xshape = [batch, height, width, in_channel]
+ wshape = [out_channel, filter_h, filter_w, in_channel]
+ oshape = xshape
+ oshape[3] = out_channel
+ ref_func = lambda dy_np, w_np, strides, padding, out_pad:
tvm.topi.testing.conv2d_transpose_nhwc_python(
+ dy_np, np.transpose(w_np, [1, 2, 3, 0]), "HWOI", strides, padding,
out_pad
+ )
+
+ dy_np = np.random.uniform(-1, 1, oshape).astype(data_dtype)
+ w_np = np.random.uniform(-1, 1, wshape).astype(data_dtype)
+
+ if data_dtype == "float16":
+ dx_np = ref_func(
+ dy_np.astype("float32"),
+ w_np.astype("float32"),
+ (stride_h, stride_w),
+ (pad_h, pad_w),
+ (0, 0),
+ )
+ dx_np = dx_np.astype("float16")
+ else:
+ dx_np = ref_func(dy_np, w_np, (stride_h, stride_w), (pad_h, pad_w),
(0, 0))
+
+ dy = te.placeholder(oshape, name="dy", dtype=data_dtype)
+ w = te.placeholder(wshape, name="dw", dtype=data_dtype)
+ dx = cudnn.conv_backward_data(
+ dy,
+ w,
+ [pad_h, pad_w],
+ [stride_h, stride_w],
+ [1, 1],
+ conv_mode=1,
+ tensor_format=tensor_format,
+ conv_dtype=conv_dtype,
+ groups=1,
+ )
+
+ s = te.create_schedule(dx.op)
+
+ dev = tvm.cuda(0)
+ f = tvm.build(s, [dy, w, dx], "cuda --host=llvm",
name="conv2d_backward_data")
+
+ dy = tvm.nd.array(dy_np, dev)
+ w = tvm.nd.array(w_np, dev)
+ dx = tvm.nd.array(dx_np, dev)
+
+ f(dy, w, dx)
+ print(np.max(np.abs(dx.numpy() - dx_np)))
+ print(np.mean(np.abs(dx.numpy() - dx_np)))
+ tvm.testing.assert_allclose(dx.numpy(), dx_np, atol=tol, rtol=tol)
+
+
[email protected]_gpu
+@requires_cudnn
+def test_conv2d_backward_data():
+ verify_conv2d_backward_data("float32", "float32", tensor_format=0,
tol=1e-5)
+ verify_conv2d_backward_data("float32", "float32", tensor_format=1,
tol=1e-2)
+ # The scipy convolve function does not support fp16, so the reference will
be computed with
+ # fp32. Use larger tolerance to be on the safe side (1e-2 also seems
mostly ok).
+ verify_conv2d_backward_data("float16", "float16", tensor_format=1,
tol=1e-1)
+
+
+def verify_conv2d_backward_filter(data_dtype, conv_dtype, tensor_format=0,
tol=1e-5):
+ batch = 3
+ in_channel = 4
+ out_channel = 16
+ filter_h, filter_w = 3, 3
+ pad_h, pad_w = 1, 1
+ stride_h, stride_w = 1, 1
+ height, width = 32, 32
+
+ if tensor_format == 0:
+ x_shape = [batch, in_channel, height, width]
+ dy_shape = [batch, out_channel, height, width]
+ else:
+ x_shape = [batch, height, width, in_channel]
+ dy_shape = [batch, height, width, out_channel]
+
+ x_np = np.random.uniform(-1, 1, x_shape).astype(data_dtype)
+ dy_np = np.random.uniform(-1, 1, dy_shape).astype(data_dtype)
+
+ dw_np = tvm.topi.testing.conv2d_backward_weight_python(
+ dy_np,
+ x_np,
+ (filter_h, filter_w),
+ (stride_h, stride_w),
+ (pad_h, pad_w),
+ "NCHW" if tensor_format == 0 else "NHWC",
+ )
+
+ x = te.placeholder(x_shape, name="x", dtype=data_dtype)
+ dy = te.placeholder(dy_shape, name="dy", dtype=data_dtype)
+ dw = cudnn.conv_backward_filter(
+ dy,
+ x,
+ (filter_h, filter_w),
+ [pad_h, pad_w],
+ [stride_h, stride_w],
+ [1, 1],
+ conv_mode=1,
+ tensor_format=tensor_format,
+ conv_dtype=conv_dtype,
+ )
+
+ s = te.create_schedule(dw.op)
+
+ dev = tvm.cuda(0)
+ f = tvm.build(s, [dy, x, dw], "cuda --host=llvm",
name="conv2d_backward_filter")
+
+ x = tvm.nd.array(x_np, dev)
+ dy = tvm.nd.array(dy_np, dev)
+ dw = tvm.nd.array(dw_np, dev)
+
+ f(dy, x, dw)
+ print(np.max(np.abs(dw.numpy() - dw_np)))
+ print(np.mean(np.abs(dw.numpy() - dw_np)))
Review comment:
ditto
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
