KellenSunderland commented on a change in pull request #12922: Support
Quantized Fully Connected by INT8 GEMM
URL: https://github.com/apache/incubator-mxnet/pull/12922#discussion_r236119445
##########
File path: src/operator/quantization/quantized_fully_connected.cc
##########
@@ -79,6 +85,145 @@ bool QuantizedFullyConnectedType(const nnvm::NodeAttrs&
attrs,
return true;
}
+bool QuantizedFullyConnectedStorageType(const nnvm::NodeAttrs& attrs,
+ const int dev_mask,
+ DispatchMode* dispatch_mode,
+ std::vector<int> *in_attrs,
+ std::vector<int> *out_attrs) {
+ *dispatch_mode = DispatchMode::kFCompute;
+ if (dev_mask == mshadow::cpu::kDevMask) {
+ *dispatch_mode = DispatchMode::kFComputeEx;
+ }
+ for (size_t i = 0; i < out_attrs->size(); i++) {
+ STORAGE_TYPE_ASSIGN_CHECK(*out_attrs, i, kDefaultStorage);
+ if (common::stype_string((*out_attrs)[i]).compare("unknown") == 0) {
+ return false;
+ }
+ }
+ for (size_t i = 0; i < in_attrs->size(); i++) {
+ STORAGE_TYPE_ASSIGN_CHECK(*in_attrs, i, kDefaultStorage);
+ if (common::stype_string((*in_attrs)[i]).compare("unknown") == 0) {
+ return false;
+ }
+ }
+ return true;
+}
+
+struct QuantizedSumInitKernelWithBias {
+ // init sum data with bias for matrix b (n)
+ MSHADOW_XINLINE static void Map(int i, int32_t *out,
+ const int8_t *bias, const float *min_out,
+ const float *max_out, const float *min_bias,
+ const float *max_bias) {
+ typedef int32_t T1;
+ typedef int8_t T2;
+ using mshadow::red::limits::MinValue;
+ using mshadow::red::limits::MaxValue;
+ float float_for_one_out_quant =
+ MaxAbs(*min_out, *max_out) / static_cast<double>(MaxValue<T1>());
+ float float_for_one_bias_quant =
+ MaxAbs(*min_bias, *max_bias) / static_cast<double>(MaxValue<T2>());
+ if (float_for_one_out_quant != 0) {
+ out[i] = bias[i] * float_for_one_bias_quant /
+ float_for_one_out_quant;
+ } else {
+ LOG(INFO) << "WARNING: QuantizedBiasAddKernel float_for_one_out_quant is
0 !";
+ out[i] = 0;
+ }
+ }
+};
+
+template<typename SrcType>
+void QuantizedFullyConnectedForward(const nnvm::NodeAttrs& attrs,
+ const OpContext &ctx,
+ const std::vector<NDArray> &in_data,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &out_data) {
+#if MSHADOW_USE_MKL == 1
+ const FullyConnectedParam& param =
nnvm::get<FullyConnectedParam>(attrs.parsed);
+ using namespace mshadow;
+ using namespace mxnet_op;
+ size_t num_inputs = param.no_bias ? 2 : 3;
+ CHECK_EQ(in_data.size(), num_inputs * 3);
+ CHECK_EQ(out_data.size(), 3U);
+ const NDArray& data = in_data[0];
+ const NDArray& weight = in_data[1];
+ const NDArray& out = out_data[0];
+ TShape dshape = data.shape();
+ TShape wshape = weight.shape();
+ TShape oshape = out.shape();
+ auto output_temp = out.data().dptr<int32_t>();
+ auto weight_temp = weight.data().dptr<SrcType>();
+ auto data_temp = data.data().dptr<SrcType>();
+ const int omp_threads =
mxnet::engine::OpenMP::Get()->GetRecommendedOMPThreadCount();
Review comment:
Tested it out here with ltrace, and strace:
https://gist.github.com/KellenSunderland/8df5ffaecb5e13c8f7e05da9201b83c3
I didn't measure the perf of the call, but it looks like it's just a library
call with no expensive calls that I can see to the OS, so this looks fine from
my POV.
----------------------------------------------------------------
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:
[email protected]
With regards,
Apache Git Services
