Skip to content

Commit c5507f3

Browse files
bo3zvloncar
bo3z
authored and
vloncar
committed
Extended support for streaming qkeras
1 parent cc874fc commit c5507f3

File tree

3 files changed

+107
-10
lines changed

3 files changed

+107
-10
lines changed

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

Lines changed: 48 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -530,6 +530,54 @@ void hard_sigmoid(stream<data_T> &data, stream<res_T> &res) {
530530
}
531531
}
532532

533+
// *************************************************
534+
// Binary TanH Activation
535+
// *************************************************
536+
template<class data_T, class res_T, typename CONFIG_T>
537+
void binary_tanh(stream<data_T> &data, stream<res_T> &res) {
538+
BinaryTanHActLoop:
539+
#pragma ii 1
540+
for (int i = 0; i < CONFIG_T::n_in / res_T::size; i++) {
541+
542+
hls_register data_T in_data = data.read();
543+
hls_register res_T out_data;
544+
545+
BinaryTanHPackLoop:
546+
#pragma unroll
547+
for (int j = 0; j < res_T::size; j++) {
548+
if(in_data[j] > 0) out_data[j] = (typename res_T::value_type) 1;
549+
else out_data[j] = (typename res_T::value_type) -1;
550+
}
551+
552+
res.write(out_data);
553+
}
554+
}
555+
556+
// *************************************************
557+
// Ternary TanH Activation
558+
// *************************************************
559+
template<class data_T, class res_T, typename CONFIG_T>
560+
void ternary_tanh(stream<data_T> &data, stream<res_T> &res) {
561+
TernaryTanHActLoop:
562+
#pragma ii 1
563+
for (int i = 0; i < CONFIG_T::n_in / res_T::size; i++) {
564+
565+
hls_register data_T in_data = data.read();
566+
hls_register res_T out_data;
567+
568+
TernaryTanHPackLoop:
569+
#pragma unroll
570+
for (int j = 0; j < res_T::size; j++) {
571+
if(in_data[j] > 1) out_data[j] = (typename res_T::value_type) 1;
572+
else if (in_data[j] <=-1) out_data[j] = (typename res_T::value_type) -1;
573+
else out_data[j] = (typename res_T::value_type) 0;
574+
}
575+
576+
res.write(out_data);
577+
}
578+
579+
}
580+
533581
}
534582

535583
#endif

‎hls4ml/templates/vivado/nnet_utils/nnet_activation_stream.h‎

Lines changed: 44 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -637,6 +637,50 @@ void prelu(hls::stream<data_T> &data, typename data_T::value_type alpha[CONFIG_T
637637
}
638638
}
639639

640+
// *************************************************
641+
// Binary TanH Activation
642+
// *************************************************
643+
template<class data_T, class res_T, typename CONFIG_T>
644+
void binary_tanh(hls::stream<data_T> &data, hls::stream<res_T> &res) {
645+
PReLUActLoop: for (int i = 0; i < CONFIG_T::n_in / res_T::size; i++) {
646+
#pragma HLS PIPELINE
647+
648+
data_T in_data = data.read();
649+
res_T out_data;
650+
#pragma HLS DATA_PACK variable=out_data
651+
652+
PReLUPackLoop: for (int j = 0; j < res_T::size; j++) {
653+
#pragma HLS UNROLL
654+
if(in_data[j] > 0) out_data[j] = (typename res_T::value_type) 1;
655+
else out_data[j] = (typename res_T::value_type) -1;
656+
}
657+
res.write(out_data);
658+
}
659+
}
660+
661+
// *************************************************
662+
// Ternary TanH Activation
663+
// *************************************************
664+
template<class data_T, class res_T, typename CONFIG_T>
665+
void ternary_tanh(hls::stream<data_T> &data, hls::stream<res_T> &res) {
666+
PReLUActLoop: for (int i = 0; i < CONFIG_T::n_in / res_T::size; i++) {
667+
#pragma HLS PIPELINE
668+
669+
data_T in_data = data.read();
670+
res_T out_data;
671+
#pragma HLS DATA_PACK variable=out_data
672+
673+
PReLUPackLoop: for (int j = 0; j < res_T::size; j++) {
674+
#pragma HLS UNROLL
675+
if(in_data[j] > 1) out_data[j] = (typename res_T::value_type) 1;
676+
else if (in_data[j] <=-1) out_data[j] = (typename res_T::value_type) -1;
677+
else out_data[j] = (typename res_T::value_type) 0;
678+
}
679+
res.write(out_data);
680+
}
681+
}
682+
683+
640684

641685
}
642686

‎test/pytest/test_qkeras.py‎

Lines changed: 15 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -113,7 +113,8 @@ def randX_100_16():
113113
#@pytest.mark.parametrize('bits', [4, 6, 8])
114114
@pytest.mark.parametrize('bits,alpha', [(4, 1), (4, 'auto_po2')])
115115
@pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
116-
def test_single_dense_activation_exact(randX_100_16, bits, alpha, backend):
116+
@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
117+
def test_single_dense_activation_exact(randX_100_16, bits, alpha, backend, io_type):
117118
'''
118119
Test a single Dense -> Activation layer topology for
119120
bit exactness with number of bits parameter
@@ -128,11 +129,12 @@ def test_single_dense_activation_exact(randX_100_16, bits, alpha, backend):
128129

129130
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(layers=['relu1'], rounding_mode='AP_RND_CONV', saturation_mode='AP_SAT')
130131
config = hls4ml.utils.config_from_keras_model(model, granularity='name')
131-
output_dir = str(test_root_path / 'hls4mlprj_qkeras_single_dense_activation_exact_{}_{}_{}'.format(bits, alpha, backend))
132+
output_dir = str(test_root_path / 'hls4mlprj_qkeras_single_dense_activation_exact_{}_{}_{}_{}'.format(bits, alpha, backend, io_type))
132133
hls_model = hls4ml.converters.convert_from_keras_model(model,
133134
hls_config=config,
134135
output_dir=output_dir,
135-
backend=backend)
136+
backend=backend,
137+
io_type=io_type)
136138
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(layers=[])
137139
hls_model.compile()
138140

@@ -168,12 +170,13 @@ def randX_100_10():
168170
(6, 10, ternary(alpha='auto'), quantized_bits(5,2), ternary(threshold=0.8), True, False),
169171
(7, 10, binary(), quantized_bits(5,2), binary(), False, True)])
170172
@pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
171-
def test_btnn(make_btnn, randX_100_10, backend):
173+
@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
174+
def test_btnn(make_btnn, randX_100_10, backend, io_type):
172175
model, is_xnor, test_no = make_btnn
173176
X = randX_100_10
174177
cfg = hls4ml.utils.config_from_keras_model(model, granularity='name')
175-
output_dir = str(test_root_path / 'hls4mlprj_btnn_{}_{}'.format(test_no, backend))
176-
hls_model = hls4ml.converters.convert_from_keras_model(model, output_dir=output_dir, hls_config=cfg, backend=backend)
178+
output_dir = str(test_root_path / 'hls4mlprj_btnn_{}_{}_{}'.format(test_no, backend, io_type))
179+
hls_model = hls4ml.converters.convert_from_keras_model(model, output_dir=output_dir, hls_config=cfg, backend=backend, io_type=io_type)
177180
hls_model.compile()
178181
y_hls = hls_model.predict(X)
179182
# hls4ml may return XNOR binary
@@ -201,7 +204,8 @@ def randX_1000_1():
201204
(quantized_relu(10)),
202205
(quantized_relu(10,5))])
203206
@pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
204-
def test_quantizer(randX_1000_1, quantizer, backend):
207+
@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
208+
def test_quantizer(randX_1000_1, quantizer, backend, io_type):
205209
'''
206210
Test a single quantizer as an Activation function.
207211
Checks the type inference through the conversion is correct without just
@@ -215,12 +219,13 @@ def test_quantizer(randX_1000_1, quantizer, backend):
215219

216220
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(layers=['quantizer'], rounding_mode='AP_RND_CONV', saturation_mode='AP_SAT')
217221
config = hls4ml.utils.config_from_keras_model(model, granularity='name')
218-
output_dir = str(test_root_path / 'hls4mlprj_qkeras_quantizer_{}_{}_{}_{}'.format(quantizer.__class__.__name__,
219-
quantizer.bits, quantizer.integer, backend))
222+
output_dir = str(test_root_path / 'hls4mlprj_qkeras_quantizer_{}_{}_{}_{}_{}'.format(quantizer.__class__.__name__,
223+
quantizer.bits, quantizer.integer, backend, io_type))
220224
hls_model = hls4ml.converters.convert_from_keras_model(model,
221225
hls_config=config,
222226
output_dir=output_dir,
223-
backend=backend)
227+
backend=backend,
228+
io_type=io_type)
224229
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(layers=[])
225230
hls_model.compile()
226231

0 commit comments

Comments
 (0)