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 2018/04/29 04:42:49 UTC
[GitHub] piiswrong closed pull request #10710: Add cuda version check to
skip building quantization ops for versions less than 8.0
piiswrong closed pull request #10710: Add cuda version check to skip building quantization ops for versions less than 8.0
URL: https://github.com/apache/incubator-mxnet/pull/10710
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/quantization/quantized_conv.cu b/src/operator/quantization/quantized_conv.cu
index 2db5416309b..a76782b8baa 100644
--- a/src/operator/quantization/quantized_conv.cu
+++ b/src/operator/quantization/quantized_conv.cu
@@ -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:
@@ -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,
@@ -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
typedef QuantizedCuDNNConvOp<int8_t, float, int32_t> QuantizedConvOpInt8;
#if DMLC_CXX11_THREAD_LOCAL
static thread_local QuantizedConvOpInt8 op;
@@ -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)
diff --git a/src/operator/quantization/quantized_fully_connected.cu b/src/operator/quantization/quantized_fully_connected.cu
index ac7ba1e21df..beecc759864 100644
--- a/src/operator/quantization/quantized_fully_connected.cu
+++ b/src/operator/quantization/quantized_fully_connected.cu
@@ -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,
@@ -49,6 +50,7 @@ struct QuantizedBiasAddKernel {
float_for_one_out_quant;
}
};
+#endif // CUDA_VERSION >= 8000
template<typename SrcType, typename DstType, typename CmpType>
void QuantizedFullyConnectedForwardGPU(const nnvm::NodeAttrs& attrs,
@@ -56,6 +58,7 @@ void QuantizedFullyConnectedForwardGPU(const nnvm::NodeAttrs& attrs,
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;
@@ -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)
diff --git a/src/operator/quantization/quantized_pooling.cu b/src/operator/quantization/quantized_pooling.cu
index 78011b885c5..2bbac5fc151 100644
--- a/src/operator/quantization/quantized_pooling.cu
+++ b/src/operator/quantization/quantized_pooling.cu
@@ -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:
@@ -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,
@@ -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
@@ -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)
----------------------------------------------------------------
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