Fix for 2D conv layers in the special case of io_parallel with full parallelization (#760)

* adding fix for fully partition io_parallel

* remove newline

* pre-commit

* update for vitis compliance

* Get rid of unroll factor and follow up in 1D

---------

Co-authored-by: Javier Duarte <jduarte@ucsd.edu>
Co-authored-by: Vladimir Loncar <vloncar@users.noreply.github.com>

Author: 3 people

hls4ml/templates/vivado/nnet_utils/nnet_conv1d_latency.h

@@ -32,7 +32,7 @@ void conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan],
 
 PartitionLoop:
     for (int i_part = 0; i_part < CONFIG_T::n_partitions; i_part++) {
-        #pragma HLS PIPELINE II=CONFIG_T::reuse_factor
+        #pragma HLS PIPELINE II=CONFIG_T::reuse_factor rewind
 
         CONFIG_T::template fill_buffer<data_T, CONFIG_T>::fill_buffer(data, data_buf, i_part);
 
@@ -45,9 +45,11 @@ void conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan],
         // Do the matrix-multiply
         Product1:
             for (int i_in = 0; i_in < mult_n_in; i_in++) {
+                #pragma HLS UNROLL
                 cache = data_buf[i_pxl][i_in];
             Product2:
                 for (int i_out = 0; i_out < mult_n_out; i_out++) {
+                    #pragma HLS UNROLL
                     mult[i_in * mult_n_out + i_out] =
                         CONFIG_T::mult_config::template product<data_T, typename CONFIG_T::mult_config::weight_t>::product(
                             cache, weights[i_in * mult_n_out + i_out]);
@@ -57,21 +59,25 @@ void conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan],
         // Initialize accumulator with input biases
         ResetAccum:
             for (int i_acc = 0; i_acc < mult_n_out; i_acc++) {
+                #pragma HLS UNROLL
                 acc[i_acc] = (typename CONFIG_T::accum_t)biases[i_acc];
             }
 
         // Accumulate multiplication result
         Accum1:
             for (int i_in = 0; i_in < mult_n_in; i_in++) {
+                #pragma HLS UNROLL
             Accum2:
                 for (int i_out = 0; i_out < mult_n_out; i_out++) {
+                    #pragma HLS UNROLL
                     acc[i_out] += mult[i_in * mult_n_out + i_out];
                 }
             }
 
         // Cast to "res_t" type
         Result:
             for (int i_res = 0; i_res < mult_n_out; i_res++) {
+                #pragma HLS UNROLL
                 *(res++) = cast<data_T, res_T, typename CONFIG_T::mult_config>(acc[i_res]);
             }
         }

hls4ml/templates/vivado/nnet_utils/nnet_conv2d_latency.h

@@ -33,7 +33,7 @@ void conv_2d_latency_cl(
 
 PartitionLoop:
     for (int i_part = 0; i_part < CONFIG_T::n_partitions; i_part++) {
-        #pragma HLS PIPELINE II=CONFIG_T::reuse_factor
+        #pragma HLS PIPELINE II=CONFIG_T::reuse_factor rewind
 
         CONFIG_T::template fill_buffer<data_T, CONFIG_T>::fill_buffer(data, data_buf, i_part);
 
@@ -46,9 +46,11 @@ void conv_2d_latency_cl(
         // Do the matrix-multiply
         Product1:
             for (int i_in = 0; i_in < mult_n_in; i_in++) {
+                #pragma HLS UNROLL
                 cache = data_buf[i_pxl][i_in];
             Product2:
                 for (int i_out = 0; i_out < mult_n_out; i_out++) {
+                    #pragma HLS UNROLL
                     mult[i_in * mult_n_out + i_out] =
                         CONFIG_T::mult_config::template product<data_T, typename CONFIG_T::mult_config::weight_t>::product(
                             cache, weights[i_in * mult_n_out + i_out]);
@@ -58,21 +60,25 @@ void conv_2d_latency_cl(
         // Initialize accumulator with input biases
         ResetAccum:
             for (int i_acc = 0; i_acc < mult_n_out; i_acc++) {
+                #pragma HLS UNROLL
                 acc[i_acc] = (typename CONFIG_T::accum_t)biases[i_acc];
             }
 
         // Accumulate multiplication result
         Accum1:
             for (int i_in = 0; i_in < mult_n_in; i_in++) {
+                #pragma HLS UNROLL
             Accum2:
                 for (int i_out = 0; i_out < mult_n_out; i_out++) {
+                    #pragma HLS UNROLL
                     acc[i_out] += mult[i_in * mult_n_out + i_out];
                 }
             }
 
         // Cast to "res_t" type
         Result:
             for (int i_res = 0; i_res < mult_n_out; i_res++) {
+                #pragma HLS UNROLL
                 *(res++) = cast<data_T, res_T, typename CONFIG_T::mult_config>(acc[i_res]);
             }
         }