piiswrong closed pull request #10538: [MXNET-318] Allow dot for fp16 on GPU
URL: https://github.com/apache/incubator-mxnet/pull/10538
This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:
As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):
diff --git a/src/operator/tensor/dot-inl.h b/src/operator/tensor/dot-inl.h
index c5f278e78a4..83571d9e4d2 100644
--- a/src/operator/tensor/dot-inl.h
+++ b/src/operator/tensor/dot-inl.h
@@ -69,9 +69,10 @@ void DotForward_(const nnvm::NodeAttrs& attrs,
<< "Binary function only support input/output with the same type";
CHECK_EQ(outputs[0].type_flag_, inputs[1].type_flag_)
<< "Binary function only support input/output with the same type";
- CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64)
- << "dot only supports float32 and float64";
- MSHADOW_SGL_DBL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
+ CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64
||
+ (outputs[0].type_flag_ == kFloat16 && ctx.run_ctx.ctx.dev_mask() ==
mshadow::gpu::kDevMask))
+ << "dot only supports float32/float64 for CPU, and
float16/float32/float64 for GPU";
+ MSHADOW_REAL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
if (inputs[0].ndim() == 1 && inputs[1].ndim() == 1) {
CHECK_NE(req[0], kAddTo) << "AddTo not yet supported";
Tensor<xpu, 1, DType> out = outputs[0].get<xpu, 1, DType>(s);
@@ -129,7 +130,7 @@ void DotBackward_(const nnvm::NodeAttrs& attrs,
Stream<xpu> *s = ctx.get_stream<xpu>();
CHECK_NE(req[0], kWriteInplace);
CHECK_NE(req[1], kWriteInplace);
- MSHADOW_SGL_DBL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
+ MSHADOW_REAL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
if (inputs[1].ndim() == 1 && inputs[2].ndim() == 1) {
Tensor<xpu, 1, DType> mout_grad = inputs[0].get<xpu, 1, DType>(s);
Tensor<xpu, 1, DType> mlhs_data = inputs[1].get<xpu, 1, DType>(s);
@@ -1126,9 +1127,10 @@ void BatchDotForward_(const nnvm::NodeAttrs& attrs,
<< "Binary function only support input/output with the same type";
CHECK_EQ(outputs[0].type_flag_, inputs[1].type_flag_)
<< "Binary function only support input/output with the same type";
- CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64)
- << "dot only supports float32 and float64";
- MSHADOW_SGL_DBL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
+ CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64
||
+ (outputs[0].type_flag_ == kFloat16 && ctx.run_ctx.ctx.dev_mask() ==
mshadow::gpu::kDevMask))
+ << "dot only supports float32/float64 for CPU, and float16/float32/float64
for GPU";
+ MSHADOW_REAL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
mshadow::Tensor<xpu, 3, DType> out = outputs[0].get<xpu, 3, DType>(s);
mshadow::Tensor<xpu, 3, DType> mlhs = inputs[0].get<xpu, 3, DType>(s);
mshadow::Tensor<xpu, 3, DType> mrhs = inputs[1].get<xpu, 3, DType>(s);
@@ -1168,9 +1170,10 @@ void BatchDotBackward_(const nnvm::NodeAttrs& attrs,
const DotParam& param = nnvm::get<DotParam>(attrs.parsed);
CHECK_NE(req[1], kWriteInplace);
CHECK_NE(req[0], kWriteInplace);
- CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64)
- << "dot only supports float32 and float64";
- MSHADOW_SGL_DBL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
+ CHECK(outputs[0].type_flag_ == kFloat32 || outputs[0].type_flag_ == kFloat64
||
+ (outputs[0].type_flag_ == kFloat16 && ctx.run_ctx.ctx.dev_mask() ==
mshadow::gpu::kDevMask))
+ << "dot only supports float32/float64 for CPU, and float16/float32/float64
for GPU";
+ MSHADOW_REAL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
mshadow::Tensor<xpu, 3, DType> mout_grad = inputs[0].get<xpu, 3, DType>(s);
mshadow::Tensor<xpu, 3, DType> mlhs_data = inputs[1].get<xpu, 3, DType>(s);
mshadow::Tensor<xpu, 3, DType> mrhs_data = inputs[2].get<xpu, 3, DType>(s);
diff --git a/tests/python/unittest/test_operator.py
b/tests/python/unittest/test_operator.py
index 5d382220a7a..3a0436e6125 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -2071,6 +2071,8 @@ def test_stn():
def test_dot():
ctx=default_context()
dtypes = ['float32', 'float64']
+ if ctx.device_type == 'gpu':
+ dtypes += ['float16']
# Test normal dot.
for data_type in dtypes:
@@ -2094,10 +2096,16 @@ def test_dot():
c = mx.sym.dot(a, b)
exe = c.simple_bind(ctx=ctx, a=a_npy.shape, b=b_npy.shape)
outputs = exe.forward(is_train=True, a=a_npy, b=b_npy)
- assert_almost_equal(outputs[0].asnumpy(), c_npy, rtol=1e-3)
- exe.backward(out_grads=[mx.nd.array(ograd_npy, mx.cpu())])
- assert_almost_equal(exe.grad_dict['a'].asnumpy(),
agrad_npy, rtol=1e-3)
- assert_almost_equal(exe.grad_dict['b'].asnumpy(),
bgrad_npy, rtol=1e-3)
+ assert_almost_equal(outputs[0].asnumpy(), c_npy,
+ rtol=1e-2 if data_type == 'float16'
else 1e-3,
+ atol=1e-2 if data_type == 'float16'
else 1e-3)
+ exe.backward(out_grads=[mx.nd.array(ograd_npy,
mx.cpu()).astype(data_type)])
+ assert_almost_equal(exe.grad_dict['a'].asnumpy(),
agrad_npy,
+ rtol=1e-2 if data_type == 'float16'
else 1e-3,
+ atol=1e-2 if data_type == 'float16'
else 1e-3)
+ assert_almost_equal(exe.grad_dict['b'].asnumpy(),
bgrad_npy,
+ rtol=1e-2 if data_type == 'float16'
else 1e-3,
+ atol=1e-2 if data_type == 'float16'
else 1e-3)
# Test dot with transpose flag using gradient checker.
def dot_sym(data_type):
@@ -2135,6 +2143,8 @@ def dot_sym_xT_yT(data_type):
@with_seed()
def test_batch_dot():
dtypes = ['float32', 'float64']
+ if default_context().device_type == 'gpu':
+ dtypes += ['float16']
for data_type in dtypes:
for batch_size in range(1, 5):
@@ -2178,16 +2188,26 @@ def test_batch_dot():
exe_add.grad_dict['a'][:] = a_init_grad_npy
exe_add.grad_dict['b'][:] = b_init_grad_npy
outputs = exe.forward(is_train=True, a=a_npy, b=b_npy)
- assert_almost_equal(outputs[0].asnumpy(), c_npy,
rtol=1e-3, atol=1e-4)
+ assert_almost_equal(outputs[0].asnumpy(), c_npy,
+ rtol=1e-2 if data_type ==
'float16' else 1e-3,
+ atol=1e-2 if data_type ==
'float16' else 1e-4)
exe.backward(out_grads=[mx.nd.array(ograd_npy,
ctx=exe._ctx)])
- assert_almost_equal(exe.grad_dict['a'].asnumpy(),
agrad_npy, rtol=1e-3, atol=1e-4)
- assert_almost_equal(exe.grad_dict['b'].asnumpy(),
bgrad_npy, rtol=1e-3, atol=1e-4)
+ assert_almost_equal(exe.grad_dict['a'].asnumpy(),
agrad_npy,
+ rtol=1e-2 if data_type ==
'float16' else 1e-3,
+ atol=1e-2 if data_type ==
'float16' else 1e-4)
+ assert_almost_equal(exe.grad_dict['b'].asnumpy(),
bgrad_npy,
+ rtol=1e-2 if data_type ==
'float16' else 1e-3,
+ atol=1e-2 if data_type ==
'float16' else 1e-4)
exe_add.forward(is_train=True, a=a_npy, b=b_npy)
exe_add.backward(out_grads=[mx.nd.array(ograd_npy,
ctx=exe._ctx)])
assert_almost_equal(exe_add.grad_dict['a'].asnumpy(),
- agrad_npy + a_init_grad_npy, rtol=1e-3, atol=1e-4)
+ agrad_npy + a_init_grad_npy,
+ rtol=1e-2 if data_type ==
'float16' else 1e-3,
+ atol=1e-2 if data_type ==
'float16' else 1e-4)
assert_almost_equal(exe_add.grad_dict['b'].asnumpy(),
- bgrad_npy + b_init_grad_npy, rtol=1e-3, atol=1e-4)
+ bgrad_npy + b_init_grad_npy,
+ rtol=1e-2 if data_type ==
'float16' else 1e-3,
+ atol=1e-2 if data_type ==
'float16' else 1e-4)
def
get_correlation(data1,data2,kernel_size,max_displacement,stride1,stride2,pad_size,is_multiply):
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
