Updated oneAPI template: io_stream kernel template.

Author: haoyanwa

hls4ml/templates/oneapi/firmware/nnet_utils/nnet_activation_stream.h

@@ -29,23 +29,33 @@ template <class data_pipe, class res_pipe, typename CONFIG_T> void linear_stream
 // *************************************************
 //       ReLU Activation
 // *************************************************
-template <class data_pipe, class res_pipe, typename CONFIG_T> void relu_stream() {
+template <class data_pipe, class res_pipe, typename CONFIG_T> 
+[[intel::use_stall_enable_clusters]] void relu_stream() {
+    using namespace nnet;
+    using ResT = typename ExtractDataType<typename ExtractPipeType<res_pipe>::value_type>::value_type;
+    [[intel::fpga_register]] typename ExtractPipeType<res_pipe>::value_type out_data;
+    
+    bool keep_going = true;
 ReLUActLoop:
-    [[intel::initiation_interval(
-        1)]] for (int i = 0; i < CONFIG_T::n_in / std::tuple_size<typename ExtractPipeType<res_pipe>::value_type>{}; i++) {
-        auto in_data = data_pipe::read();
-        typename ExtractPipeType<res_pipe>::value_type out_data;
+    [[intel::initiation_interval(1)]]
+    while(keep_going) {
+        for (int i = 0; i < CONFIG_T::n_in / std::tuple_size<ResT>{}; i++) {
+            [[intel::fpga_register]] auto in_data = data_pipe::read();
+ReLUPackLoop:
+            #pragma unroll
+            for (int j = 0; j < std::tuple_size<ResT>{}; j++) {
+                if (in_data.data[j] > 0)
+                    out_data.data[j] = in_data.data[j];
+                else
+                    out_data.data[j] = 0;
+            }
 
-    ReLUPackLoop:
-        #pragma unroll
-        for (int j = 0; j < std::tuple_size<typename ExtractPipeType<res_pipe>::value_type>{}; j++) {
-            if (in_data[j] > 0)
-                out_data[j] = in_data[j];
-            else
-                out_data[j] = 0;
-        }
+            out_data.sop = in_data.sop;
+            out_data.eop = in_data.eop;
+            res_pipe::write(out_data);
 
-        res_pipe::write(out_data);
+            keep_going = !in_data.eop;
+        }
     }
 }
 

hls4ml/templates/oneapi/firmware/nnet_utils/nnet_dense_stream.h

@@ -7,15 +7,34 @@
 
 namespace nnet {
 
-// Note:  DataPack logic removed, at least in the initial version
+// Restartable streaming kernel implementation.
+// Computation is carried out in a while-1 loop as long as there is valid input.
+// The loop breaks when the end-of-packet signal is asserted by upstream task.
 template <class data_pipe, class res_pipe, typename CONFIG_T>
-void dense_resource_stream(typename CONFIG_T::weight_t weights, typename CONFIG_T::bias_t biases) {
+[[intel::use_stall_enable_clusters]] void dense_resource_stream(const typename CONFIG_T::weight_t weights, const typename CONFIG_T::bias_t biases) {
+    using namespace nnet;
+    using DataT = typename ExtractDataType<typename ExtractPipeType<data_pipe>::value_type>::value_type;
+    using ResT = typename ExtractDataType<typename ExtractPipeType<res_pipe>::value_type>::value_type;
+    
+    [[intel::fpga_register]] typename ExtractPipeType<res_pipe>::value_type resbeat;
 
-    [[intel::fpga_register]] typename ExtractPipeType<res_pipe>::value_type res;
-    [[intel::fpga_register]] auto data = data_pipe::read();
-    dense_resource<typename ExtractPipeType<data_pipe>::value_type, typename ExtractPipeType<res_pipe>::value_type,
-                   CONFIG_T>(data, res, weights, biases);
-    res_pipe::write(res);
+    bool keep_going = true;
+    bool did_read_input;
+    [[intel::initiation_interval(1)]]
+    while (keep_going) {
+        did_read_input = false;
+        [[intel::fpga_register]] auto databeat = data_pipe::read(did_read_input);
+
+        if (did_read_input) {
+            dense_resource<DataT, ResT, CONFIG_T>(databeat.data, resbeat.data, weights, biases);
+
+            resbeat.sop = databeat.sop;
+            resbeat.eop = databeat.eop;
+
+            res_pipe::write(resbeat);
+            keep_going = !databeat.eop;
+        }
+    }
 }
 
 } // namespace nnet