Add general transpose for vivado/vitis

hls4ml/backends/vivado/passes/reshaping_templates.py

@@ -1,3 +1,7 @@
+from math import prod
+
+import numpy as np
+
 from hls4ml.backends.template import FunctionCallTemplate, LayerConfigTemplate
 from hls4ml.model.layers import Resize, Transpose, ZeroPadding1D, ZeroPadding2D
 
@@ -97,16 +101,64 @@ def format(self, node):
 
 # Transpose templates
 
-transpose_config_template = """struct config{index} : nnet::transpose_config {{
-    static const unsigned depth = {depth};
-    static const unsigned height = {height};
-    static const unsigned width = {width};
-    static constexpr unsigned perm[3] = {{{perm_str}}};
-}};\n"""
 
-transpose_function_template = 'nnet::transpose_{dim}<{input_t}, {output_t}, {config}>({input}, {output});'
-
-transpose_include_list = ['nnet_utils/nnet_array.h', 'nnet_utils/nnet_stream.h']
+transpose_include_list = ['nnet_utils/nnet_transpose.h', 'nnet_utils/nnet_transpose_stream.h']
+
+transpose_config_template = """struct {config_name} {{
+    static const unsigned dims = {dims};
+    static const unsigned N = {N};
+    static const unsigned* const from_shape;
+    static const unsigned* const to_shape;
+    static const unsigned* const perm;
+    static const unsigned* const perm_strides;
+}};
+
+unsigned {config_name}_from_shape[{dims}] = {{{from_shape}}};
+unsigned {config_name}_to_shape[{dims}] = {{{to_shape}}};
+unsigned {config_name}_perm[{dims}] = {{{perm}}};
+unsigned {config_name}_perm_strides[{dims}] = {{{perm_strides}}};
+
+const unsigned* const {config_name}::from_shape = {config_name}_from_shape;
+const unsigned* const {config_name}::to_shape = {config_name}_to_shape;
+const unsigned* const {config_name}::perm = {config_name}_perm;
+const unsigned* const {config_name}::perm_strides = {config_name}_perm_strides;
+"""
+
+transpose_function_template = 'nnet::transpose<{input_t}, {output_t}, {config_name}>({input}, {output});'
+
+
+def permute_config_gen(name: str, shape: tuple[int, ...], perm: tuple[int, ...]):
+    """
+    Generate a configuration string for a permute operation. Operates by mapping the output index to input input index by:
+     - unravel the output index
+     - map each dimension to the corresponding stride in the input tensor, sum
+    The operation can be expressed as:
+
+    new_shape = tuple(shape[i] for i in perm)
+    strides = np.cumprod((shapes[1:] + (1,))[::-1])[::-1]
+    perm_strides = [strides[i] for i in perm]
+    out[index] = inp[np.dot(np.unravel_index(index, new_shape), perm_strides)]
+
+    Args:
+        name (str): The name of the configuration.
+        shape (tuple[int, ...]): The shape of the input tensor.
+        perm (tuple[int, ...]): The permutation of the dimensions.
+
+    Returns:
+        str: The formatted configuration string for the permute operation.
+    """
+    new_shape = tuple(shape[i] for i in perm)
+    strides = np.cumprod((shape[1:] + (1,))[::-1])[::-1]
+    perm_strides = tuple(int(strides[i]) for i in perm)
+    return transpose_config_template.format(
+        dims=len(shape),
+        N=prod(shape),
+        from_shape=', '.join(str(x) for x in shape),
+        perm=', '.join(str(x) for x in perm),
+        perm_strides=', '.join(str(x) for x in perm_strides),
+        to_shape=', '.join(str(x) for x in new_shape),
+        config_name=name,
+    )
 
 
 class TransposeConfigTemplate(LayerConfigTemplate):
@@ -115,18 +167,18 @@ def __init__(self):
         self.template = transpose_config_template
 
     def format(self, node):
-        params = self._default_config_params(node)
-
-        return self.template.format(**params)
+        shape = tuple(node.get_input_variable().shape)
+        perm = tuple(node.get_attr('perm'))
+        name = f'config{node.index}'
+        return permute_config_gen(name, shape, perm)
 
 
 class TransposeFunctionTemplate(FunctionCallTemplate):
     def __init__(self):
-        super().__init__(Transpose, include_header=transpose_include_list)
         self.template = transpose_function_template
+        super().__init__(Transpose, include_header=transpose_include_list)
 
     def format(self, node):
         params = self._default_function_params(node)
-        params['dim'] = node.get_attr('dim')
-
+        params['config_name'] = f'config{node.index}'
         return self.template.format(**params)

hls4ml/model/layers.py

@@ -1221,8 +1221,7 @@ def initialize(self):
         perm = self.get_attr('perm')
         self.set_attr('dim', f'{len(inp.shape)}d')
 
-        if len(perm) > 3:
-            raise Exception('ERROR: Transpose of tensors with rank > 3 is not yet supported.')
+        # TODO: dim>3 is only supported for vivado/vitis backend
 
         # ONNX double transpose specific, sometimes ONNX injects
         # useless double transpose layers when converting
@@ -1242,11 +1241,14 @@ def initialize(self):
             self.set_attr('depth', 1)
             self.set_attr('height', inp.shape[0])
             self.set_attr('width', inp.shape[1])
-        elif len(shape) > 2:
+        elif len(shape) == 3:
             dims = [f'OUT_DEPTH_{self.index}', f'OUT_HEIGHT_{self.index}', f'OUT_WIDTH_{self.index}']
             self.set_attr('depth', inp.shape[0])
             self.set_attr('height', inp.shape[1])
             self.set_attr('width', inp.shape[2])
+        elif len(shape) > 3:
+            # Differentiate between 2/3/3+ dim does not really appear to be needed. To be removed?
+            dims = [f'OUT_DIM_{i}_{self.index}' for i in range(1, len(shape) + 1)]
         self.add_output_variable(shape, dims, precision=inp.type.precision)
 
 

hls4ml/model/optimizer/passes/convert_to_channels_last.py

@@ -12,14 +12,13 @@ class ChannelsLastConverter(OptimizerPass):
     and adding a transpose layer for the inputs and outputs, if necessary'''
 
     def match(self, node):
+        # If this parameter has not been set, this model does not need to be converted
+        if 'ChannelsLastConversion' not in node.model.config.config['HLSConfig']['Model']:
+            return False  # No littering of unused property
         if not hasattr(node, 'channels_last_converted'):
             return True
 
     def transform(self, model, node):
-        # If this parameter has not been set, this model does not need to be converted
-        if 'ChannelsLastConversion' not in model.config.config['HLSConfig']['Model']:
-            node.channels_last_converted = True
-            return False
         outshape = node.get_output_variable().shape
 
         if isinstance(node, Input):
@@ -103,11 +102,6 @@ def transform(self, model, node):
                 input = previous_node.name
                 outshape = previous_node.get_output_variable().shape
 
-                if (model.config.config['IOType'] == 'io_stream') and len(outshape) == 3:
-                    raise Exception(
-                        'No 3D transpose available in io_stream, this model cannot be converted to channels-last'
-                    )
-
                 if len(outshape) == 2:
                     attributes = {'perm': [1, 0]}
                 else:

hls4ml/templates/vivado/nnet_utils/nnet_stream.h

@@ -179,29 +179,6 @@ void broadcast_stream(hls::stream<data_T> &data, hls::stream<res_T> &res) {
     }
 }
 
-template <class data_T, class res_T, typename CONFIG_T>
-void transpose_2d(hls::stream<data_T> &data, hls::stream<res_T> &res) {
-    typename data_T::value_type data_array[CONFIG_T::height * CONFIG_T::width];
-    #pragma HLS ARRAY_PARTITION variable=data_array complete
-
-    for (int i = 0; i < CONFIG_T::height * CONFIG_T::width / data_T::size; i++) {
-        #pragma HLS PIPELINE
-        data_T in_data = data.read();
-        for (int j = 0; j < data_T::size; j++) {
-            data_array[i * data_T::size + j] = typename data_T::value_type(in_data[j]);
-        }
-    }
-
-    for (int i = 0; i < CONFIG_T::height * CONFIG_T::width / res_T::size; i++) {
-        #pragma HLS PIPELINE
-        res_T out_data;
-        PRAGMA_DATA_PACK(out_data)
-        for (int j = 0; j < res_T::size; j++) {
-            out_data[j] = typename res_T::value_type(data_array[j * data_T::size + i]);
-        }
-        res.write(out_data);
-    }
-}
 } // namespace nnet
 
 #endif

hls4ml/templates/vivado/nnet_utils/nnet_transpose.h

@@ -0,0 +1,39 @@
+#ifndef NNET_PERMUTE_H_
+#define NNET_PERMUTE_H_
+
+namespace nnet {
+
+struct transpose_config {
+    static const unsigned dims;
+    static const unsigned N;
+    // vivado/vitis hls can't index constexpr array for some reason
+    // and vivado hls don't like template recursion either (vitis is fine)
+    // thus this appears to be the only workaround (or overkill it with codegen)
+    static const unsigned *const from_shape;
+    static const unsigned *const to_shape;
+    static const unsigned *const perm;
+    static const unsigned *const perm_strides;
+};
+
+template <typename CONFIG_T> unsigned transfer_idx(int index) {
+    // Given output idx in c-order flat array, return input idx
+    int idx = 0;
+    for (int i = CONFIG_T::dims - 1; i >= 0; i--) {
+        idx += (index % CONFIG_T::to_shape[i]) * CONFIG_T::perm_strides[i];
+        index /= CONFIG_T::to_shape[i];
+    }
+    return idx;
+}
+
+template <typename data_T, typename res_T, typename CONFIG_T>
+void transpose(const data_T data[CONFIG_T::N], res_T res[CONFIG_T::N]) {
+    for (int i = 0; i < CONFIG_T::N; i++) {
+        #pragma HLS UNROLL
+        int idx = transfer_idx<CONFIG_T>(i);
+        res[i] = data[idx];
+    }
+}
+
+} // namespace nnet
+
+#endif

hls4ml/templates/vivado/nnet_utils/nnet_transpose_stream.h

@@ -0,0 +1,67 @@
+#ifndef NNET_TRANSPOSE_STREAM_H
+#define NNET_TRANSPOSE_STREAM_H
+
+#include "hls_stream.h"
+#include "nnet_transpose.h"
+#include <type_traits>
+
+namespace nnet {
+
+template <typename data_T, typename res_T, typename CONFIG_T>
+typename std::enable_if<CONFIG_T::dims == 2, void>::type transpose(hls::stream<data_T> &data, hls::stream<res_T> &res) {
+    // #pragma HLS INLINE RECURSIVE
+    typename data_T::value_type data_array[CONFIG_T::N];
+    #pragma HLS ARRAY_PARTITION variable=data_array complete
+
+    for (int i = 0; i < CONFIG_T::N / data_T::size; i++) {
+        #pragma HLS PIPELINE
+        data_T in_data = data.read();
+        for (int j = 0; j < data_T::size; j++) {
+            #pragma HLS UNROLL
+            data_array[i * data_T::size + j] = typename data_T::value_type(in_data[j]);
+        }
+    }
+
+    for (int i = 0; i < CONFIG_T::N / res_T::size; i++) {
+        #pragma HLS PIPELINE
+        res_T out_data;
+        PRAGMA_DATA_PACK(out_data)
+        for (int j = 0; j < res_T::size; j++) {
+            #pragma HLS UNROLL
+            out_data[j] = typename res_T::value_type(data_array[j * CONFIG_T::from_shape[1] + i]);
+        }
+        res.write(out_data);
+    }
+}
+
+// This sfinae is for vivado_hls, which has some overhead using the transfer_idx in io_stream.
+// In vitis both performs exactly the same, thus this is not removed out of convenience.
+template <typename data_T, typename res_T, typename CONFIG_T>
+typename std::enable_if<CONFIG_T::dims != 2, void>::type transpose(hls::stream<data_T> &data, hls::stream<res_T> &res) {
+    // #pragma HLS INLINE RECURSIVE
+    typename data_T::value_type data_array[CONFIG_T::N];
+    #pragma HLS ARRAY_PARTITION variable=data_array complete
+
+    for (int i = 0; i < CONFIG_T::N / data_T::size; i++) {
+        #pragma HLS PIPELINE
+        data_T in_data = data.read();
+        for (int j = 0; j < data_T::size; j++) {
+            #pragma HLS UNROLL
+            data_array[i * data_T::size + j] = typename data_T::value_type(in_data[j]);
+        }
+    }
+
+    for (int i = 0; i < CONFIG_T::N / res_T::size; i++) {
+        #pragma HLS PIPELINE
+        res_T out_data;
+        PRAGMA_DATA_PACK(out_data)
+        for (int j = 0; j < res_T::size; j++) {
+            #pragma HLS UNROLL
+            out_data[j] = typename res_T::value_type(data_array[transfer_idx<CONFIG_T>(i * res_T::size + j)]);
+        }
+        res.write(out_data);
+    }
+}
+
+} // namespace nnet
+#endif

test/pytest/test_transpose_concat.py

@@ -54,3 +54,37 @@ def test_accuracy(data, keras_model, hls_model):
     y_hls4ml = hls_model.predict(X).reshape(y_keras.shape)
     # "accuracy" of hls4ml predictions vs keras
     np.testing.assert_allclose(y_keras, y_hls4ml, rtol=0, atol=1e-04, verbose=True)
+
+
+@pytest.fixture(scope='module')
+def keras_model_highdim():
+    inp = Input(shape=(2, 3, 4, 5, 6), name='input_1')
+    out = Permute((3, 5, 4, 1, 2))(inp)
+    model = Model(inputs=inp, outputs=out)
+    return model
+
+
+@pytest.fixture(scope='module')
+def data_highdim():
+    X = np.random.randint(-128, 127, (100, 2, 3, 4, 5, 6)) / 128
+    X = X.astype(np.float32)
+    return X
+
+
+@pytest.mark.parametrize('io_type', ['io_stream', 'io_parallel'])
+@pytest.mark.parametrize('backend', ['Vivado', 'Vitis'])
+def test_highdim_permute(data_highdim, keras_model_highdim, io_type, backend):
+    X = data_highdim
+    model = keras_model_highdim
+
+    model_hls = hls4ml.converters.convert_from_keras_model(
+        model,
+        io_type=io_type,
+        backend=backend,
+        output_dir=str(test_root_path / f'hls4mlprj_highdim_transpose_{backend}_{io_type}'),
+    )
+    model_hls.compile()
+    y_keras = model.predict(X)
+    y_hls4ml = model_hls.predict(X).reshape(y_keras.shape)  # type: ignore
+
+    assert np.all(y_keras == y_hls4ml)