Skip to content

Parallel conv partial fix #1380

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 12 commits into
base: main
Choose a base branch
from
Open

Parallel conv partial fix #1380

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
f7535b7
pointwise conv 1d partial fix
calad0i Sep 8, 2025
e88b856
pointwise inherit all property
calad0i Sep 8, 2025
7a0ee88
inline for parallel conv, pointwise only for npart=1 (otherwise II wi…
calad0i Sep 8, 2025
d077b72
avoid unexpected shared state
calad0i Sep 8, 2025
f349e41
rm useless warning
calad0i Sep 9, 2025
39d416b
rm line
calad0i Sep 9, 2025
7a6f1db
.
calad0i Sep 9, 2025
343ae5c
Merge branch 'main' into conv-partial-fix
jmitrevs Sep 22, 2025
e530e18
fix pragma
calad0i Sep 23, 2025
43b8cee
Merge branch 'main' into conv-partial-fix
JanFSchulte Oct 3, 2025
8143d75
vivado unsupported pragma
calad0i Oct 6, 2025
4aaf11c
disable also for vitis (2025.1 crashes...)
calad0i Oct 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion hls4ml/backends/catapult/passes/pointwise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ def match(self, node):

def transform(self, model, node):
dim = node.__class__.__name__[-2:] # '1D' or '2D'
new_attrs = {k: v for k, v in node.attributes.items() if k not in ('trace', 'precision', 'reuse_factor')}
new_attrs = node.attributes.attributes.copy()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I mentioned this in a comment in the conversation. I think trace, precision, and reuse_factor are regenerated no matter what, so the values you copy here get overriden (unless something has changed from before). It may be an hls4ml behavior worth revisiting and potentially revising, but I don't think this change fixes anything.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reuse_factor defined under Model key is not propagated as expected otherwise. Rm'ed warning if the update opr is trivial

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Did you see why it's not propagated properly? Shouldn't it come from a configuration in the first place?

pw_node = model.make_node(
'PointwiseConv' + dim, node.name, new_attrs, node.inputs.copy(), outputs=node.outputs.copy()
)
Expand Down
2 changes: 1 addition & 1 deletion hls4ml/backends/oneapi/passes/pointwise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,7 +147,7 @@ def match(self, node):

def transform(self, model, node):
dim = node.__class__.__name__[-2:] # '1D' or '2D'
new_attrs = {k: v for k, v in node.attributes.items() if k not in ('trace', 'precision', 'reuse_factor')}
new_attrs = node.attributes.attributes.copy()
pw_node = model.make_node(
'PointwiseConv' + dim, node.name, new_attrs, node.inputs.copy(), outputs=node.outputs.copy()
)
Expand Down
2 changes: 1 addition & 1 deletion hls4ml/backends/quartus/passes/pointwise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ def match(self, node):

def transform(self, model, node):
dim = node.__class__.__name__[-2:] # '1D' or '2D'
new_attrs = {k: v for k, v in node.attributes.items() if k not in ('trace', 'precision', 'reuse_factor')}
new_attrs = node.attributes.attributes.copy()
pw_node = model.make_node(
'PointwiseConv' + dim, node.name, new_attrs, node.inputs.copy(), outputs=node.outputs.copy()
)
Expand Down
18 changes: 12 additions & 6 deletions hls4ml/backends/vivado/passes/convolution_templates.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,8 +101,11 @@ def format(self, node):
and node.get_attr('strategy').lower() == 'latency'
and node.model.config.get_config_value('IOType') == 'io_parallel'
)
if is_pointwise_parallel_latency:
params['conv_fn'] = f'{namespace}::pointwise_conv_{node.index}'

n_partitions = node.attributes['n_partitions']

if is_pointwise_parallel_latency and n_partitions == 1:
params['conv_fn'] = 'nnet::BatchedDenseForConv1D'
else:
if node.get_attr('strategy').lower() == 'latency':
params['conv_fn'] = 'nnet::Conv1DLatency'
Expand All @@ -115,11 +118,14 @@ def format(self, node):
conv_config = self.template.format(**params)

mult_params = self._default_config_params(node)
if is_pointwise_parallel_latency:
mult_params['n_in'] = int(
node.get_attr('in_width') * node.get_attr('n_chan') * node.get_attr('filt_width') / mult_params['reuse']
if is_pointwise_parallel_latency and n_partitions == 1:
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Doesn't this restrict the whole function to a single (fully parallel) case?

Also, ff we ever fix the HLS code to make it work for any combination of PF/RF, will this line be the only change we need to make on the python side? If so, maybe document it

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This does restrict it to the full parallel case, since otherwise the performance was worse than the standard implementation. Will need to check again with the last pragma change...

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought that operation instead of function would actually cause the synthesis to fail, not the pragma being ignored

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In vivado_hls it fails, but in vitis it just ignored the pragma

mult_params['n_in'] = (
node.get_attr('in_width')
* node.get_attr('n_chan')
* node.get_attr('filt_width')
// mult_params['n_partitions']
)
mult_params['n_out'] = int(node.get_attr('in_width') * node.get_attr('n_filt') / mult_params['reuse'])
mult_params['n_out'] = node.get_attr('in_width') * node.get_attr('n_filt') // mult_params['n_partitions']
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Are n_partitions and mult_params['n_partions'] different?

else:
mult_params['n_in'] = node.get_attr('n_chan') * node.get_attr('filt_width')
mult_params['n_out'] = node.get_attr('n_filt')
Expand Down
2 changes: 1 addition & 1 deletion hls4ml/backends/vivado/passes/pointwise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,7 @@ def match(self, node):
def transform(self, model, node):
dim = node.__class__.__name__[-2:] # '1D' or '2D'
# to remove warning, since these get set again
new_attrs = {k: v for k, v in node.attributes.items() if k not in ('trace', 'precision', 'reuse_factor')}
new_attrs = node.attributes.attributes.copy()
pw_node = model.make_node(
'PointwiseConv' + dim, node.name, new_attrs, node.inputs.copy(), outputs=node.outputs.copy()
)
Expand Down
84 changes: 0 additions & 84 deletions hls4ml/backends/vivado/passes/pointwise_codegen.py
View file
Open in desktop

This file was deleted.

3 changes: 1 addition & 2 deletions hls4ml/backends/vivado/vivado_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,7 +183,6 @@ def _register_flows(self):
'vivado:generate_conv_streaming_instructions',
'vivado:apply_resource_strategy',
'vivado:generate_conv_im2col',
'vivado:generate_pointwise_conv1_d',
'vivado:generate_unrolled_dense_resource',
'vivado:set_pipeline_style',
'vivado:d_a_latency_dense_template',
Expand Down Expand Up @@ -418,7 +417,7 @@ def init_conv1d(self, layer):
if user_pf is not None and layer_pf is not None:
if user_pf != layer_pf:
warn(
f'For layer {layer.name}, parallelization factor of {layer_pf} is defined in the proxy-model, but is overridden by the user to {user_pf}.' # noqa: E501
f'Parallelization factor of {layer_pf} is embedded in layer {layer.name} is overridden by user config to {user_pf}.' # noqa: E501
)

valid_pf = self.get_valid_conv_partition_splits(1, out_width)
Expand Down
58 changes: 54 additions & 4 deletions hls4ml/templates/vitis/nnet_utils/nnet_conv1d.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,13 @@ void conv_1d_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan], res_T res[CO
typename CONFIG_T::weight_t weights[CONFIG_T::filt_width * CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
// Inlining helps reduce latency, but may also cause timing issues in some cases, use carefully.
//#pragma HLS INLINE recursive
// But without inlining Vitis HLS doesn't respect the parallelization factor config ¯\_(ツ)_/
// Vitis2025.1 hangs in RTL simulation with this, though

#pragma HLS INLINE recursive

// #pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
// ↑ This makes II=2 in for all n_partitions > 1, no matter what the actual II should be

CONFIG_T::template conv_kernel<data_T, res_T, CONFIG_T>::conv(data, res, weights, biases);
}
Expand All @@ -50,7 +56,13 @@ void pointwise_conv_1d_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan],
assert(CONFIG_T::filt_width == 1);

// Inlining helps reduce latency, but may also cause timing issues in some cases, use carefully.
//#pragma HLS INLINE recursive
// But without inlining Vitis HLS doesn't respect the parallelization factor config ¯\_(ツ)_/¯
// #pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
Copy link
Contributor

@jmitrevs jmitrevs Sep 8, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe erase commented out pragmas throughout?


#pragma HLS INLINE recursive

// #pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
// ↑ This makes II=2 in for all n_partitions > 1, no matter what the actual II should be

CONFIG_T::template conv_kernel<data_T, res_T, CONFIG_T>::conv(data, res, weights, biases);
}
Expand All @@ -61,7 +73,7 @@ class Conv1DLatency : public nnet::Conv1DKernel<data_T, res_T, CONFIG_T> {
static void conv(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan], res_T res[CONFIG_T::out_width * CONFIG_T::n_filt],
typename CONFIG_T::weight_t weights[CONFIG_T::filt_width * CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
//#pragma HLS INLINE region
// #pragma HLS INLINE recursive
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is already disabled, and now remains disabled but with a different action. Why?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Was enbaled and disabled when changing things around. Though, since inline region doesn't exist in vitis, the current one seems to be making a bit more sense...

conv_1d_latency_cl<data_T, res_T, CONFIG_T>(data, res, weights, biases);
}
};
Expand All @@ -72,11 +84,49 @@ class Conv1DResource : public nnet::Conv1DKernel<data_T, res_T, CONFIG_T> {
static void conv(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan], res_T res[CONFIG_T::out_width * CONFIG_T::n_filt],
typename CONFIG_T::weight_t weights[CONFIG_T::filt_width * CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
//#pragma HLS INLINE region
// #pragma HLS INLINE recursive
conv_1d_resource_cl<data_T, res_T, CONFIG_T>(data, res, weights, biases);
}
};

template <class data_T, class res_T, typename CONFIG_T>
class BatchedDenseForConv1D : public nnet::Conv1DKernel<data_T, res_T, CONFIG_T> {
Copy link
Contributor

@jmitrevs jmitrevs Sep 8, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we add a comment to say the purpose of this code (and also for the 1D version)?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The PR allows modification from contributors. Feel free to add some if you find it necessary.

public:
static void conv(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan], res_T res[CONFIG_T::out_width * CONFIG_T::n_filt],
typename CONFIG_T::weight_t weights[CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {

#pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
#pragma HLS INLINE RECURSIVE
data_T data_tmp[CONFIG_T::n_partitions][CONFIG_T::in_width * CONFIG_T::n_chan / CONFIG_T::n_partitions];
#pragma HLS ARRAY_PARTITION variable=data_tmp complete dim=0
res_T res_tmp[CONFIG_T::n_partitions][CONFIG_T::out_width * CONFIG_T::n_filt / CONFIG_T::n_partitions];
#pragma HLS ARRAY_PARTITION variable=res_tmp complete dim=0

for (int jj = 0; jj < CONFIG_T::n_partitions; jj++) {
#pragma HLS UNROLL
for (int ii = 0; ii < CONFIG_T::in_width * CONFIG_T::n_chan / CONFIG_T::n_partitions; ii++) {
#pragma HLS UNROLL
data_tmp[jj][ii] = data[jj * CONFIG_T::in_width * CONFIG_T::n_chan / CONFIG_T::n_partitions + ii];
}
}

#pragma HLS ALLOCATION operation instances=nnet::pointwise_conv_1d_latency_cl<data_T, res_T, CONFIG_T> limit=1

for (int jj = 0; jj < CONFIG_T::n_partitions; jj++) {
nnet::pointwise_conv_1d_latency_cl<data_T, res_T, CONFIG_T>(data_tmp[jj], res_tmp[jj], weights, biases);
}

for (int jj = 0; jj < CONFIG_T::n_partitions; jj++) {
#pragma HLS UNROLL
for (int ii = 0; ii < CONFIG_T::out_width * CONFIG_T::n_filt / CONFIG_T::n_partitions; ii++) {
#pragma HLS UNROLL
res[jj * CONFIG_T::out_width * CONFIG_T::n_filt / CONFIG_T::n_partitions + ii] = res_tmp[jj][ii];
}
}
}
};

} // namespace nnet

#endif
16 changes: 8 additions & 8 deletions hls4ml/templates/vitis/nnet_utils/nnet_conv1d_latency.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,14 +86,14 @@ void conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan],
}

template <class data_T, class res_T, typename CONFIG_T>
void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan / CONFIG_T::reuse_factor],
res_T res[CONFIG_T::out_width * CONFIG_T::n_filt / CONFIG_T::reuse_factor],
void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_chan / CONFIG_T::n_partitions],
res_T res[CONFIG_T::out_width * CONFIG_T::n_filt / CONFIG_T::n_partitions],
typename CONFIG_T::weight_t weights[CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
assert(CONFIG_T::filt_width == 1);

typename CONFIG_T::accum_t mult[CONFIG_T::out_width * CONFIG_T::n_filt * CONFIG_T::n_chan / CONFIG_T::reuse_factor];
typename CONFIG_T::accum_t acc[CONFIG_T::out_width / CONFIG_T::reuse_factor][CONFIG_T::n_filt];
typename CONFIG_T::accum_t mult[CONFIG_T::out_width * CONFIG_T::n_filt * CONFIG_T::n_chan / CONFIG_T::n_partitions];
typename CONFIG_T::accum_t acc[CONFIG_T::out_width / CONFIG_T::n_partitions][CONFIG_T::n_filt];

#pragma HLS ARRAY_PARTITION variable=mult complete dim=0
#pragma HLS ARRAY_PARTITION variable=acc complete dim=0
Expand All @@ -111,7 +111,7 @@ void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_c

// Convolve, saving all multiplication results to accumulate later
ConvOut:
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::reuse_factor; ii++) {
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::n_partitions; ii++) {
ConvFilt:
for (int ff = 0; ff < CONFIG_T::n_filt; ff++) {
ConvChan:
Expand All @@ -133,7 +133,7 @@ void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_c
} // end output loop

// Initialize accumulator with input biases
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::reuse_factor; ii++) {
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::n_partitions; ii++) {
for (int ff = 0; ff < CONFIG_T::n_filt; ff++) {
#pragma HLS UNROLL
acc[ii][ff] = biases[ff];
Expand All @@ -142,7 +142,7 @@ void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_c

// Accumulate multiplication result
AccumOut:
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::reuse_factor; ii++) {
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::n_partitions; ii++) {
AccumFilt:
for (int ff = 0; ff < CONFIG_T::n_filt; ff++) {
// Do "dot product" sum within filter and sum over channels
Expand All @@ -155,7 +155,7 @@ void pointwise_conv_1d_latency_cl(data_T data[CONFIG_T::in_width * CONFIG_T::n_c
} // end output loop

// Cast to "res_t" type
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::reuse_factor; ii++) {
for (int ii = 0; ii < CONFIG_T::out_width / CONFIG_T::n_partitions; ii++) {
for (int ff = 0; ff < CONFIG_T::n_filt; ff++) {
#pragma HLS UNROLL
res[ii * CONFIG_T::n_filt + ff] = cast<data_T, res_T, typename CONFIG_T::mult_config>(acc[ii][ff]);
Expand Down
17 changes: 14 additions & 3 deletions hls4ml/templates/vitis/nnet_utils/nnet_conv2d.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,13 @@ void conv_2d_cl(
typename CONFIG_T::weight_t weights[CONFIG_T::filt_height * CONFIG_T::filt_width * CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
// Inlining helps reduce latency, but may also cause timing issues in some cases, use carefully.
//#pragma HLS INLINE recursive
// But without inlining Vitis HLS doesn't respect the parallelization factor config ¯\_(ツ)_/
// Vitis2025.1 hangs in RTL simulation with this, though

#pragma HLS INLINE recursive

// #pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
// ↑ This makes II=2 in for all n_partitions > 1, no matter what the actual II should be

if (CONFIG_T::strategy == nnet::latency || CONFIG_T::strategy == nnet::distributed_arithmetic) {
conv_2d_latency_cl<data_T, res_T, CONFIG_T>(data, res, weights, biases);
Expand All @@ -60,9 +66,14 @@ void pointwise_conv_2d_cl(data_T data[CONFIG_T::in_height * CONFIG_T::in_width *
typename CONFIG_T::weight_t weights[CONFIG_T::n_chan * CONFIG_T::n_filt],
typename CONFIG_T::bias_t biases[CONFIG_T::n_filt]) {
assert(CONFIG_T::filt_width == 1);

// Inlining helps reduce latency, but may also cause timing issues in some cases, use carefully.
//#pragma HLS INLINE recursive
// But without inlining Vitis HLS doesn't respect the parallelization factor config ¯\_(ツ)_/
// Vitis2025.1 hangs in RTL simulation with this, though

#pragma HLS INLINE recursive

// #pragma HLS PIPELINE II = CONFIG_T::reuse_factor * CONFIG_T::n_partitions
// ↑ This makes II=2 in for all n_partitions > 1, no matter what the actual II should be

// Nothing special to be done for io_parallel implementation
if (CONFIG_T::strategy == nnet::latency || CONFIG_T::strategy == nnet::distributed_arithmetic) {
Expand Down
Loading
Loading