You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2018/08/08 07:11:24 UTC
[GitHub] eric-haibin-lin closed pull request #11886: Improve error message
of cudnn operators
eric-haibin-lin closed pull request #11886: Improve error message of cudnn operators
URL: https://github.com/apache/incubator-mxnet/pull/11886
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/nn/cudnn/cudnn_convolution-inl.h b/src/operator/nn/cudnn/cudnn_convolution-inl.h
index 4b1cbbe7057..aa38c18a73c 100644
--- a/src/operator/nn/cudnn/cudnn_convolution-inl.h
+++ b/src/operator/nn/cudnn/cudnn_convolution-inl.h
@@ -685,6 +685,7 @@ class CuDNNConvolutionOp {
const int kMaxAlgos = 10;
int nalgo = kMaxAlgos;
int i = 0;
+ size_t min_memory_needs = 0;
// Forward Algorithm Find/Get, v6 and earlier
if (CUDNN_MAJOR == 6 && param_.layout.value() == mshadow::kNHWC) {
// In cuDNNv6, for kNHWC, only CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM is
@@ -715,10 +716,16 @@ class CuDNNConvolutionOp {
while (i < nalgo
&& (fwd_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == conv::kLimited
- && fwd_algo[i].memory > workspace_byte)))
+ && fwd_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs =
+ (i == 0) ? fwd_algo[i].memory : std::min(min_memory_needs, fwd_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a forward convolution algorithm.";
+ LOG(FATAL) << nalgo << " forward algorithms with minimum memory requirement "
+ << min_memory_needs << " bytes have been tried. Workspace size is set to "
+ << workspace_byte << " bytes, please consider reducing the batch/model size, "
+ << "or increasing workspace size.";
} else {
forward_algo_.Set(fwd_algo[i].algo, false);
}
@@ -749,10 +756,17 @@ class CuDNNConvolutionOp {
while (i < nalgo
&& (bwd_filter_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == conv::kLimited
- && bwd_filter_algo[i].memory > workspace_byte)))
+ && bwd_filter_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs = (i == 0) ?
+ bwd_filter_algo[i].memory :
+ std::min(min_memory_needs, bwd_filter_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a backward filter convolution algorithm.";
+ LOG(FATAL) << nalgo << " backward filter algorithms with minimum memory requirement "
+ << min_memory_needs << " bytes have been tried. Workspace size is set to "
+ << workspace_byte << " bytes, please consider reducing the batch/model size, "
+ << "or increasing workspace size.";
} else {
back_algo_w_.Set(bwd_filter_algo[i].algo, false);
}
@@ -783,10 +797,17 @@ class CuDNNConvolutionOp {
while (i < nalgo
&& (bwd_data_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == conv::kLimited
- && bwd_data_algo[i].memory > workspace_byte)))
+ && bwd_data_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs = (i == 0) ?
+ bwd_data_algo[i].memory :
+ std::min(min_memory_needs, bwd_data_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a backward data convolution algorithm.";
+ LOG(FATAL) << nalgo << " backward data algorithms with minimum memory requirement "
+ << min_memory_needs << " bytes have been tried. Workspace size is set to "
+ << workspace_byte << " bytes, please consider reducing the batch/model size, "
+ << "or increasing workspace size.";
} else {
back_algo_.Set(bwd_data_algo[i].algo, false);
}
@@ -833,7 +854,9 @@ class CuDNNConvolutionOp {
}
}
auto mode = param_.cudnn_tune.value() == conv::kOff ? " get " : " find ";
- LOG(FATAL) << "Failed to" << mode << "any " << kernel_name << " convolution algorithm.";
+ LOG(FATAL) << "Failed to" << mode << "any " << kernel_name << " convolution algorithm. "
+ << " with workspace size of " << workspace_byte << " bytes,"
+ << " please consider reducing batch/model size or increasing the workspace size";
}
void GetTempSize(const OpContext& ctx) {
diff --git a/src/operator/nn/cudnn/cudnn_deconvolution-inl.h b/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
index cb0de4c961b..74baab8f3b2 100644
--- a/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
+++ b/src/operator/nn/cudnn/cudnn_deconvolution-inl.h
@@ -618,6 +618,7 @@ class CuDNNDeconvolutionOp {
const int kMaxAlgos = 10;
int nalgo = kMaxAlgos;
int i = 0;
+ size_t min_memory_needs = 0;
// Forward Algorithm Find/Get, v6 and earlier
if (CUDNN_MAJOR == 6 && param_.layout.value() == mshadow::kNHWC) {
// In cuDNNv6, for kNHWC, only CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM is
@@ -648,11 +649,19 @@ class CuDNNDeconvolutionOp {
while (i < nalgo
&& (fwd_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == deconv::kLimited
- && fwd_algo[i].memory > workspace_byte)))
+ && fwd_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs = (i == 0) ?
+ fwd_algo[i].memory :
+ std::min(min_memory_needs, fwd_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a 'forward' convolution algorithm " <<
- "(for use in deconvolution operator backprop-to-data).";
+ LOG(FATAL) << nalgo << " forward algorithms"
+ << " (for use in deconvolution operator backprop-to-data)"
+ << " with minimum memory requirement " << min_memory_needs
+ << " bytes have been tried. Workspace size is set to " << workspace_byte
+ << " bytes, please consider reducing the batch/model size,"
+ << " or increasing workspace size.";
} else {
forward_algo_.Set(fwd_algo[i].algo, false);
}
@@ -683,11 +692,19 @@ class CuDNNDeconvolutionOp {
while (i < nalgo
&& (bwd_filter_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == deconv::kLimited
- && bwd_filter_algo[i].memory > workspace_byte)))
+ && bwd_filter_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs = (i == 0) ?
+ bwd_filter_algo[i].memory :
+ std::min(min_memory_needs, bwd_filter_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a backward filter convolution algorithm " <<
- "(for use in deconvolution operator backprop-to-filter).";
+ LOG(FATAL) << nalgo << " backward filter algorithms"
+ << " (for use in deconvolution operator backprop-to-filter)"
+ << " with minimum memory requirement " << min_memory_needs
+ << " bytes have been tried. Workspace size is set to " << workspace_byte
+ << " bytes, please consider reducing the batch/model size,"
+ << " or increasing workspace size.";
} else {
back_algo_w_.Set(bwd_filter_algo[i].algo, false);
}
@@ -718,11 +735,19 @@ class CuDNNDeconvolutionOp {
while (i < nalgo
&& (bwd_data_algo[i].status != CUDNN_STATUS_SUCCESS
|| (param_.cudnn_tune.value() == deconv::kLimited
- && bwd_data_algo[i].memory > workspace_byte)))
+ && bwd_data_algo[i].memory > workspace_byte))) {
++i;
+ min_memory_needs = (i == 0) ?
+ bwd_data_algo[i].memory :
+ std::min(min_memory_needs, bwd_data_algo[i].memory);
+ }
if (i == nalgo) {
- LOG(FATAL) << "Failed to find a backward data convolution algorithm." <<
- "(for use in deconvolution operator forward inference).";
+ LOG(FATAL) << nalgo << " backward data algorithms"
+ << " (for use in deconvolution operator forward inference) with"
+ << " minimum memory requirement " << min_memory_needs
+ << " bytes have been tried. Workspace size is set to " << workspace_byte
+ << " bytes, please consider reducing the batch/model size,"
+ << " or increasing workspace size.";
} else {
back_algo_.Set(bwd_data_algo[i].algo, false);
}
@@ -774,7 +799,9 @@ class CuDNNDeconvolutionOp {
}
}
auto mode = param_.cudnn_tune.value() == conv::kOff ? " get " : " find ";
- LOG(FATAL) << "Failed to" << mode << "any " << kernel_name << " deconvolution algorithm.";
+ LOG(FATAL) << "Failed to" << mode << "any " << kernel_name << " deconvolution algorithm"
+ << " with workspace size of " << workspace_byte << " bytes,"
+ << " please consider reducing batch/model size or increasing the workspace size";
}
void GetTempSize(const OpContext& ctx) {
----------------------------------------------------------------
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:
users@infra.apache.org
With regards,
Apache Git Services