Merge pull request #655 from bo3z/sofstign-stream-opt

Quartus Streaming Softsign (PR #585 contd.)
fastmachinelearning · Sep 20, 2022 · 9ca7af2 · 9ca7af2
2 parents 40ae7f9 + 9271648
commit 9ca7af2
Show file tree

Hide file tree

Showing 2 changed files with 27 additions and 36 deletions.
diff --git a/hls4ml/templates/quartus/firmware/nnet_utils/nnet_activation_stream.h b/hls4ml/templates/quartus/firmware/nnet_utils/nnet_activation_stream.h
@@ -223,6 +223,8 @@ template<class data_T, class res_T, typename CONFIG_T>
 void softsign(stream<data_T> &data, stream<res_T> &res) {
     #include "activation_tables/softsign_table.tb"
 
+    static const int MAX_VALUE = 8;
+
     SoftsignActLoop:
     #pragma ii 1
     for (int i = 0; i < CONFIG_T::n_in / res_T::size; i++) {
@@ -232,11 +234,21 @@ void softsign(stream<data_T> &data, stream<res_T> &res) {
         SoftsignPackLoop:
         #pragma unroll
         for (int j = 0; j < res_T::size; j++) {
-            hls_register int data_round = (in_data[j]*CONFIG_T::table_size/16).to_int();
-            hls_register int index = data_round + 8*CONFIG_T::table_size/16;
-            if (index < 0) index = 0;
-            else if (index > CONFIG_T::table_size-1) index = CONFIG_T::table_size-1;
-            out_data[j] = softsign_table[index];
+            hls_register typename data_T::value_type absValue;;
+            if(in_data[j] < 0){
+                absValue = -in_data[j];
+            }
+            else{
+                absValue = in_data[j];
+            }
+            ac_int<16> index = (absValue * CONFIG_T::table_size / MAX_VALUE).to_int();
+            if (absValue > MAX_VALUE) index = CONFIG_T::table_size - 1;
+            if(in_data[j] < 0) {
+                out_data[j] = -(typename res_T::value_type) softsign_table[index];
+            }
+            else {
+                out_data[j] = (typename res_T::value_type) softsign_table[index];
+            }
         }
 
         res.write(out_data);

diff --git a/test/pytest/test_softsign.py b/test/pytest/test_softsign.py
@@ -7,48 +7,27 @@
 
 test_root_path = Path(__file__).parent
 
-def flat_distribution(shape):
-    return np.random.rand(*shape)
-
-
-@pytest.fixture()
-def generate_data(function, input_shape):
-    return function((1000, *input_shape))
-
-
-# TODO: include latency strategy with flat_distribution when it can be made to pass
-@pytest.mark.parametrize('backend,strategy,function,input_shape,io_type', [   
-                            ('Vivado', 'stable', flat_distribution, (4,), 'io_parallel'),
-                            ('Quartus', 'stable', flat_distribution, (4,), 'io_parallel'),
-
-                            # IO_stram avaliable just for VIVADO
-                            ('Vivado', 'stable', flat_distribution, (4,), 'io_stream'),
-                            ('Vivado', 'stable', flat_distribution, (4, 4, 3), 'io_stream')
+@pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
+@pytest.mark.parametrize('input_shape, io_type', [
+                            ((8, ), 'io_parallel'),
+                            ((8, ), 'io_stream'),
+                            ((8, 8, 3), 'io_stream')
                         ])
-def test_softsign(backend, strategy, generate_data, input_shape, io_type):
-    X = generate_data
+def test_softsign(backend, input_shape, io_type):
+    X = np.random.rand(1000, *input_shape)
     model = tf.keras.models.Sequential()
     model.add(tf.keras.layers.Activation(input_shape=input_shape, activation='softsign', name='softsign'))
     model.compile()
 
-    f_type = 'ac_fixed<18,8,true,AC_RND,AC_SAT>' if backend == 'Quartus' else 'ap_fixed<18,8,AP_RND,AP_SAT>'
-    cfg = hls4ml.utils.config_from_keras_model(model, granularity='name')
-    cfg['LayerName']['softsign']['Strategy'] = strategy
-    cfg['LayerName']['softsign']['inv_table_t'] = f_type
-    cfg['LayerName']['softsign']['exp_table_t'] = f_type
-
-    odir = str(test_root_path / 'hls4mlprj_softsign_{}'.format(strategy))
+    cfg = hls4ml.utils.config_from_keras_model(model, granularity='name')    
+    odir = str(test_root_path / 'hls4mlprj_softsign_{}_{}_{}'.format(backend, io_type, str(input_shape)))
     hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=cfg, io_type=io_type,
                                                            output_dir=odir, backend=backend)
     hls_model.compile()
 
     y_keras = model.predict(X)
     y_hls4ml = hls_model.predict(X).reshape(y_keras.shape)
-
     acc_hls4ml = accuracy_score(np.argmax(y_keras, axis=-1).ravel(), np.argmax(y_hls4ml, axis=-1).ravel())
 
     print('Accuracy hls4ml relative to keras: {}'.format(acc_hls4ml))
-
-    assert acc_hls4ml >= 0.98
-
-
+    assert acc_hls4ml >= 0.97