diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 5d7806701e6664b860167175448001a3d0b54a0a..fcd6f9d788d0af1cad6de5259e5e181e76ac96bc 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -15,7 +15,7 @@ gecho () {
# the repos themselves are cloned in the Dockerfile
BREVITAS_COMMIT=989cdfdba4700fdd900ba0b25a820591d561c21a
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=1893584c83dc4500fd92733d500e80903bab1d5d
+HLSLIB_COMMIT=13e9b0772a27a3a1efc40c878d8e78ed09efb716
PYVERILATOR_COMMIT=1d89cb0d4e0c97469cc6352c611f876ec13edfa6
PYNQSHELL_COMMIT=0c82a61b0ec1a07fa275a14146233824ded7a13d
diff --git a/run-docker.sh b/run-docker.sh
index 186efc322a8f437be0371b5a142a9dd524d1abf3..e07556716db335421f57a390f1e6a17168ac058b 100755
--- a/run-docker.sh
+++ b/run-docker.sh
@@ -96,7 +96,7 @@ gecho "Port-forwarding for Netron $NETRON_PORT:$NETRON_PORT"
gecho "Vivado IP cache dir is at $VIVADO_IP_CACHE"
gecho "Using default PYNQ board $PYNQ_BOARD"
-DOCKER_INTERACTIVE = ""
+DOCKER_INTERACTIVE=""
if [ "$1" = "test" ]; then
gecho "Running test suite (all tests)"
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
index 03119b3fbea12cf9065e561089ca5875a8f622b0..3e40ad70208909551365c51324153859ccc79ceb 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
@@ -41,10 +41,19 @@ from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
# output 0 is the output tensor, shape NHWC:
# = (1, OFMDim, OFMDim, (ConvKernelDim^2)*IFMChannels)
+# note: the actual data layout produced by the hlslib kernels is different
+# for depthwise and non-depthwise ops.
+# * non-depthwise SWG: (1, OFMDim, OFMDim, K, K, IFMChannels/SIMD, SIMD)
+# * depthwise SWG: (1, OFMDim, OFMDim, IFMChannels/SIMD, K, K, SIMD)
+# see test_fpgadataflow_slidingwindow.py for an example of how to transform
+# between the two layouts
+
class ConvolutionInputGenerator(HLSCustomOp):
- """Class that corresponds to finn-hlslib ConvolutionInputGenerator
- (sliding window) function."""
+ """Class that corresponds to one of the finn-hlslib ConvolutionInputGenerator
+ (sliding window) function variants. Depending on the combination of
+ attributes (e.g. depthwise or not, whether k % stride is 0) a different
+ variant will be picked for the actual HLS implementation."""
def __init__(self, onnx_node):
super().__init__(onnx_node)
@@ -60,6 +69,7 @@ class ConvolutionInputGenerator(HLSCustomOp):
# FINN DataTypes for inputs, weights, outputs
"inputDataType": ("s", True, ""),
"outputDataType": ("s", True, ""),
+ "depthwise": ("i", False, 0),
# FPGA resource type for ConvolutionInputGenerator input buffer
# auto -- let Vivado HLS decide
# block -- use BRAM
@@ -106,7 +116,6 @@ class ConvolutionInputGenerator(HLSCustomOp):
pad = 0
ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad)
assert ifm_ch % simd == 0, "SIMD must divide IFMChannels"
- assert k % stride == 0, "stride must divide kernel size k"
wf = int((k * k * ifm_ch) // simd)
folded_oshape = (1, ofm_dim, ofm_dim, wf, simd)
return folded_oshape
@@ -313,12 +322,27 @@ class ConvolutionInputGenerator(HLSCustomOp):
"ultra": "ap_resource_uram()",
}
hls_ram_style = map_to_hls_ram_style[ram_style]
- self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<ConvKernelDim1, IFMChannels1, Input_precision1, IFMDim1,
- OFMDim1, SIMD1, Stride1> (in0, out, numReps, {});""".format(
- node.op_type, hls_ram_style
- )
- ]
+ hls_call = node.op_type
+ # check if non optimized ConvolutionInputGenerator is needed
+ k = self.get_nodeattr("ConvKernelDim")
+ stride = self.get_nodeattr("Stride")
+ if k % stride != 0:
+ hls_call += "_kernel_stride"
+
+ if self.get_nodeattr("depthwise") == 1:
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}_dws<ConvKernelDim1, IFMChannels1, Input_precision1, IFMDim1,
+ OFMDim1, SIMD1, Stride1> (in0, out, numReps, {});""".format(
+ hls_call, hls_ram_style
+ )
+ ]
+ else:
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}<ConvKernelDim1, IFMChannels1, Input_precision1, IFMDim1,
+ OFMDim1, SIMD1, Stride1> (in0, out, numReps, {});""".format(
+ hls_call, hls_ram_style
+ )
+ ]
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
diff --git a/src/finn/custom_op/im2col.py b/src/finn/custom_op/im2col.py
index 16446c15d46ee7996162f864708f7fde6cfedaf3..82a6b140f7af1be4e5c0f429d077b99c7865383e 100644
--- a/src/finn/custom_op/im2col.py
+++ b/src/finn/custom_op/im2col.py
@@ -21,8 +21,6 @@ def get_im2col_indices_nchw(
"""Returns im2col indices."""
# First figure out what the size of the output should be
N, C, H, W = x_shape
- assert (H + 2 * padding - field_height) % stride_y == 0
- assert (W + 2 * padding - field_width) % stride_x == 0
out_height = compute_conv_output_dim(H, field_height, stride_y, padding)
out_width = compute_conv_output_dim(W, field_width, stride_x, padding)
@@ -70,6 +68,9 @@ def im2col_indices_nchw(
# * ifm is the number of input channels
# * k is the convolutional kernel size
+# note: for the innermost (dot product) dimension of k*k*ifm, we
+# assume an internal ordering (k, k, ifm)
+
class Im2Col(CustomOp):
def get_nodeattr_types(self):
diff --git a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
index 5051bf34dc690daf8b6186859d3717cc8e217eee..b5fc85caf274edc9e7afc52df962862fa8a99ba3 100644
--- a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
+++ b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
@@ -78,7 +78,7 @@ def make_single_im2col_modelwrapper(k, ifm_ch, ifm_dim, ofm_dim, simd, stride, i
def make_single_slidingwindow_modelwrapper(
- k, ifm_ch, ifm_dim, ofm_dim, simd, stride, idt
+ k, ifm_ch, ifm_dim, ofm_dim, simd, stride, idt, dw=0
):
odt = idt
inp = helper.make_tensor_value_info(
@@ -102,6 +102,7 @@ def make_single_slidingwindow_modelwrapper(
Stride=stride,
inputDataType=idt.name,
outputDataType=odt.name,
+ depthwise=dw,
)
graph = helper.make_graph(
nodes=[SlidingWindow_node],
@@ -126,25 +127,29 @@ def prepare_inputs(input_tensor):
# input datatype
@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT2])
# kernel size
-@pytest.mark.parametrize("k", [2, 4])
+@pytest.mark.parametrize("k", [2, 3])
# input dimension
-@pytest.mark.parametrize("ifm_dim", [4, 6, 8])
+@pytest.mark.parametrize("ifm_dim", [6, 8])
# input channels
-@pytest.mark.parametrize("ifm_ch", [2, 4]) # , 2, 3, 4])
+@pytest.mark.parametrize("ifm_ch", [2, 4])
# Stride
@pytest.mark.parametrize("stride", [1, 2])
# execution mode
@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
# input channel parallelism ("SIMD")
@pytest.mark.parametrize("simd", [1, 2])
+# depthwise
+@pytest.mark.parametrize("dw", [0, 1])
@pytest.mark.slow
@pytest.mark.vivado
-def test_fpgadataflow_slidingwindow(idt, k, ifm_dim, ifm_ch, stride, exec_mode, simd):
+def test_fpgadataflow_slidingwindow(
+ idt, k, ifm_dim, ifm_ch, stride, exec_mode, simd, dw
+):
ofm_dim = int(((ifm_dim - k) / stride) + 1)
x = gen_finn_dt_tensor(idt, (1, ifm_dim, ifm_dim, ifm_ch))
model = make_single_slidingwindow_modelwrapper(
- k, ifm_ch, ifm_dim, ofm_dim, simd, stride, idt
+ k, ifm_ch, ifm_dim, ofm_dim, simd, stride, idt, dw
)
if exec_mode == "cppsim":
@@ -168,6 +173,12 @@ def test_fpgadataflow_slidingwindow(idt, k, ifm_dim, ifm_ch, stride, exec_mode,
k, ifm_ch, ifm_dim, ofm_dim, simd, stride, idt
)
y_expected = oxe.execute_onnx(golden, input_dict)["outp"]
- # if idt == DataType.BIPOLAR:
- # y_expected = 2 * y_expected - 1
- assert (y_produced == y_expected).all()
+ if dw == 0:
+ assert (y_produced == y_expected).all()
+ else:
+ y_expected = y_expected.reshape(
+ 1, ofm_dim, ofm_dim, k * k, ifm_ch // simd, simd
+ )
+ y_expected = y_expected.transpose(0, 1, 2, 4, 3, 5)
+ y_expected = y_expected.reshape(1, ofm_dim, ofm_dim, ifm_ch * k * k)
+ assert (y_produced == y_expected).all()