diff --git a/docker/Dockerfile.finn b/docker/Dockerfile.finn
index e42aebbe489c4e381c08f02905326eb12134291f..84078aec59856305062b9c0f50d17a2450fe3bbf 100644
--- a/docker/Dockerfile.finn
+++ b/docker/Dockerfile.finn
@@ -97,7 +97,7 @@ ARG FINN_EXP_COMMIT="af6102769226b82b639f243dc36f065340991513"
ARG BREVITAS_COMMIT="a5b71d6de1389d3e7db898fef72e014842670f03"
ARG PYVERILATOR_COMMIT="0c3eb9343500fc1352a02c020a736c8c2db47e8e"
ARG CNPY_COMMIT="4e8810b1a8637695171ed346ce68f6984e585ef4"
-ARG HLSLIB_COMMIT="f1eecde2d894a6d8971555924c0df147dc5ba033"
+ARG HLSLIB_COMMIT="da7b47cd65a967b76554a0dda74c097803c5e550"
ARG OMX_COMMIT="1dfc4aa2f2895632742cd5751520c6b472feb74e"
ARG AVNET_BDF_COMMIT="2d49cfc25766f07792c0b314489f21fe916b639b"
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
index 8d401c40dc3efbfef5c387bdfc2f777a2cfb553b..97ec89daae388109d78c1d5b21aa8239e2f0dc9d 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
@@ -264,7 +264,7 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
dilation_h, dilation_w = dilation
# since mmv != 1 is not supported yet, we set mmv for now to 1
- # mmv = 1
+ mmv = 1
# see https://github.com/Xilinx/finn-hlslib/blob/master/slidingwindow.h
if self.use_parallel_window_output():
exp_cycles = k_w + ofm_dim_w
@@ -272,10 +272,21 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
cycles_read_block = ifm_dim_w * ifm_ch / simd
cycles_write_block = ofm_dim_w * k_w * ifm_ch / simd
exp_cycles = cycles_read_block + cycles_write_block
+ elif self.get_nodeattr("depthwise") == 1:
+ if stride_h > 1 or stride_w > 1:
+ cycles_write_block = (ofm_dim_w * k_w * k_h * (ifm_ch / simd)) / mmv
+ cycles_read_block = stride_w * ifm_dim_w * (ifm_ch / simd)
+ max_cycles = max(cycles_write_block, cycles_read_block)
+ exp_cycles = (
+ ifm_dim_w * k_h * dilation_h * (ifm_ch / simd)
+ + ofm_dim_h * max_cycles
+ )
+ else:
+ cycles_read_block = ifm_ch / simd * (k_w - 1) - (k_w - 1)
+ cycles_write_block = ofm_dim_w * k_w * ifm_ch / simd
+ exp_cycles = cycles_read_block + cycles_write_block
else:
- cycles_read_block = ifm_ch / simd * (k_w - 1) - (k_w - 1)
- cycles_write_block = ofm_dim_w * k_w * ifm_ch / simd
- exp_cycles = cycles_read_block + cycles_write_block
+ exp_cycles = 1 + ofm_dim_w * k_w * ifm_ch / simd
return int(exp_cycles)
@@ -561,6 +572,15 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
"ultra": "ap_resource_uram()",
}
hls_ram_style = map_to_hls_ram_style[ram_style]
+ (
+ ifm_ch,
+ ifm_dim,
+ ofm_dim,
+ k,
+ stride,
+ dilation,
+ ) = self.get_1d_conv_attrs_normalized()
+ stride_x = np.prod(stride)
# check which ConvolutionInputGenerator is needed
if self.use_parallel_window_output():
@@ -586,20 +606,34 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
)
]
else:
- hls_call = "ConvolutionInputGenerator_1D_dws_lowbuffer"
- self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<ConvKernelDim1_x, IFMChannels1,
- Input_precision1, IFMDim1_x, OFMDim1_x,
- SIMD1> (in0, out, numReps, {});""".format(
- hls_call, hls_ram_style
- )
- ]
+ if stride_x > 1:
+ # temporarily use old ConvolutionInputGenerator_NonSquare_dws
+ # for depthwise with stride > 1
+ # note that both x and y stride are set to same (hlslib bug)
+ hls_call = "ConvolutionInputGenerator_NonSquare_dws"
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}<ConvKernelDim1_x, 1, IFMChannels1,
+ Input_precision1, IFMDim1_x, 1, OFMDim1_x, 1,
+ SIMD1, Stride1_x, Stride1_x
+ > (in0, out, numReps, {});""".format(
+ hls_call, hls_ram_style
+ )
+ ]
+ else:
+ hls_call = "ConvolutionInputGenerator_1D_dws_lowbuffer"
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}<ConvKernelDim1_x, IFMChannels1,
+ Input_precision1, IFMDim1_x, OFMDim1_x,
+ SIMD1> (in0, out, numReps, {});""".format(
+ hls_call, hls_ram_style
+ )
+ ]
else:
hls_call = "ConvolutionInputGenerator_1D_lowbuffer"
self.code_gen_dict["$DOCOMPUTE$"] = [
"""{}<ConvKernelDim1_x, IFMChannels1,
Input_precision1, IFMDim1_x, OFMDim1_x,
- SIMD1> (in0, out, numReps, {});""".format(
+ Stride1_x, SIMD1> (in0, out, numReps, {});""".format(
hls_call, hls_ram_style
)
]
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index f7cb0a16b9d3893066eaeaae416ea0259ab6a915..6cd555c52c17e86a7d0e0da7040322636c951f7f 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -213,6 +213,10 @@ class InferConvInpGen(Transformation):
1D depthwise separable convolutions"""
% n.name
)
+ if stride_h > 1 or stride_w > 1:
+ assert (
+ stride_h < k_h and stride_w < k_w
+ ), """%s: Stride value must be smaller than kernel dim"""
ConvInpGen_node = helper.make_node(
"ConvolutionInputGenerator1D",
[ConvInpGen_input],
diff --git a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
index f97b247dd654a1ab393b0181571c63465053491b..27e1907508f94e6d65ac57d313a9b3e9dd824f5a 100644
--- a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
+++ b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
@@ -156,8 +156,7 @@ def prepare_inputs(input_tensor):
# input channels
@pytest.mark.parametrize("ifm_ch", [1, 4])
# Stride
-# @pytest.mark.parametrize("stride", [[1, 1], [2, 1]])
-@pytest.mark.parametrize("stride", [[1, 1]])
+@pytest.mark.parametrize("stride", [[1, 1], [2, 1]])
# Dilation
@pytest.mark.parametrize("dilation", [[1, 1], [2, 1]])
# execution mode