Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Add cuda version check to skip building quantization ops for versions less than 8.0 #10710

Merged
merged 3 commits into from
Apr 29, 2018
Merged

Add cuda version check to skip building quantization ops for versions less than 8.0 #10710

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7526dda
Add cuda version check to skip building quantization ops for version …
reminisce Apr 26, 2018
363de1e
Clearer error message on cuda and cudnn versions
reminisce Apr 28, 2018
bac2783
Trigger CI
anirudh2290 Apr 28, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 6 additions & 5 deletions src/operator/quantization/quantized_conv.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ struct QuantizedBiasAddKernel {
}
};

#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
template<typename SrcType, typename DstType, typename CmpType>
class QuantizedCuDNNConvOp {
public:
Expand Down Expand Up @@ -260,7 +260,7 @@ class QuantizedCuDNNConvOp {
float alpha_ = 1.0f;
float beta_ = 0.0f;
}; // class QuantizedCuDNNConvOp
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000

void QuantizedConvForwardGPU(const nnvm::NodeAttrs& attrs,
const OpContext& ctx,
Expand All @@ -270,7 +270,7 @@ void QuantizedConvForwardGPU(const nnvm::NodeAttrs& attrs,
const ConvolutionParam& param = nnvm::get<ConvolutionParam>(attrs.parsed);
CHECK_EQ(param.kernel.ndim(), 2U)
<< "QuantizedConvForward<gpu> only supports 2D convolution for now";
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this need to have an else clause? What is the expected behavior if quantizedpoolforwadgpu is called with cuda version < 8 ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There will be an error message printed saying that otherwise is not supported. I have made it clearer on specifying cudnn and cuda versions.

typedef QuantizedCuDNNConvOp<int8_t, float, int32_t> QuantizedConvOpInt8;
#if DMLC_CXX11_THREAD_LOCAL
static thread_local QuantizedConvOpInt8 op;
Expand All @@ -280,8 +280,9 @@ void QuantizedConvForwardGPU(const nnvm::NodeAttrs& attrs,
op.Init(param, ctx, {inputs[0].shape_, inputs[1].shape_}, {outputs[0].shape_});
op.Forward(ctx, inputs, req, outputs);
#else
LOG(FATAL) << "QuantizedConvForward<gpu> only supports cudnnConvolutionForward for now";
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
LOG(FATAL) << "QuantizedConvForward<gpu> only supports cudnnConvolutionForward "
"with CUDNN >= 6.0 and CUDA >= 8.0";
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
}

NNVM_REGISTER_OP(_contrib_quantized_conv)
Expand Down
6 changes: 6 additions & 0 deletions src/operator/quantization/quantized_fully_connected.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
namespace mxnet {
namespace op {

#if CUDA_VERSION >= 8000
// value + bias_value * (range1 / limit_range1) * (limit_range2 / range2)
struct QuantizedBiasAddKernel {
MSHADOW_XINLINE static void Map(int i, size_t k, int32_t *out,
Expand All @@ -49,13 +50,15 @@ struct QuantizedBiasAddKernel {
float_for_one_out_quant;
}
};
#endif // CUDA_VERSION >= 8000

template<typename SrcType, typename DstType, typename CmpType>
void QuantizedFullyConnectedForwardGPU(const nnvm::NodeAttrs& attrs,
const OpContext &ctx,
const std::vector<TBlob> &inputs,
const std::vector<OpReqType> &req,
const std::vector<TBlob> &outputs) {
#if CUDA_VERSION >= 8000
const FullyConnectedParam& param = nnvm::get<FullyConnectedParam>(attrs.parsed);
using namespace mshadow;
using namespace mxnet_op;
Expand Down Expand Up @@ -113,6 +116,9 @@ void QuantizedFullyConnectedForwardGPU(const nnvm::NodeAttrs& attrs,
outputs[1].dptr<float>(), outputs[2].dptr<float>(),
inputs[7].dptr<float>(), inputs[8].dptr<float>());
}
#else
LOG(FATAL) << "QuantizedFullyConnectedForwardGPU only supports CUDA >= 8.0";
#endif // CUDA_VERSION >= 8000
}

NNVM_REGISTER_OP(_contrib_quantized_fully_connected)
Expand Down
11 changes: 6 additions & 5 deletions src/operator/quantization/quantized_pooling.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@
namespace mxnet {
namespace op {

#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
template<typename DType>
class QuantizedCuDNNPoolingOp {
public:
Expand Down Expand Up @@ -115,7 +115,7 @@ class QuantizedCuDNNPoolingOp {
cudnnTensorDescriptor_t out_desc_;
cudnnPoolingDescriptor_t pool_desc_;
}; // class QuantizedCuDNNPoolingOp
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000

void QuantizedPoolingForwardGPU(const nnvm::NodeAttrs& attrs,
const OpContext& ctx,
Expand All @@ -125,7 +125,7 @@ void QuantizedPoolingForwardGPU(const nnvm::NodeAttrs& attrs,
const PoolingParam& param = nnvm::get<PoolingParam>(attrs.parsed);
CHECK_EQ(param.kernel.ndim(), 2U)
<< "QuantizedPoolingForward<gpu> only supports 2D convolution for now";
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
#if MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
#if DMLC_CXX11_THREAD_LOCAL
static thread_local QuantizedCuDNNPoolingOp<int8_t> op;
#else
Expand All @@ -134,8 +134,9 @@ void QuantizedPoolingForwardGPU(const nnvm::NodeAttrs& attrs,
op.Init(param, {inputs[0].shape_}, {outputs[0].shape_});
op.Forward(ctx.get_stream<gpu>(), inputs, req, outputs);
#else
LOG(FATAL) << "QuantizedPoolingForward<gpu> only supports cudnnPoolingForward for now";
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6
LOG(FATAL) << "QuantizedPoolingForward<gpu> only supports cudnnPoolingForward "
"with CUDNN >= 6.0 and CUDA >= 8.0";
#endif // MXNET_USE_CUDNN == 1 && CUDNN_MAJOR >= 6 && CUDA_VERSION >= 8000
}

NNVM_REGISTER_OP(_contrib_quantized_pooling)
Expand Down