diff --git a/src/finn/custom_op/fpgadataflow/fmpadding_batch.py b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
index 27dfab54ec6d483d948dd383e54a44117d7c1a65..99f959bf59f06d6ab5b71dd7245d657f4964cca4 100644
--- a/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
@@ -17,9 +17,17 @@ class FMPadding_Batch(HLSCustomOp):
def get_nodeattr_types(self):
my_attrs = {
# spatial size of input images
- "ImgDim": ("i", True, 0),
+ "ImgDim": ("ints", True, []), # [H, W] = [Y, X]
# total padding (per dimension) to apply
- "Padding": ("i", True, 2),
+ # NOTE: Current padding scheme that is applied tries to pad the same
+ # amount of zeros in front and behind the image for each dimension.
+ # As an example, a padding scheme such as [1, x, 3, x] is equal
+ # to [2, x, 2, x]
+ "Padding": (
+ "ints",
+ True,
+ [1, 1, 1, 1],
+ ), # [H_begin, W_begin, H_end, W_end] = [Y_begin, X_begin, Y_end, X_end]
# number of channels in input image
"NumChannels": ("i", True, 0),
# SIMD Input parallelism
@@ -38,31 +46,33 @@ class FMPadding_Batch(HLSCustomOp):
def get_padded_odim(self):
"Return the padded spatial size of the output."
-
- idim = self.get_nodeattr("ImgDim")
+ idim_h, idim_w = self.get_nodeattr("ImgDim")
pad = self.get_nodeattr("Padding")
- return idim + pad
+ pad_h = pad[0] + pad[2]
+ pad_w = pad[1] + pad[3]
+ odim_h = idim_h + pad_h
+ odim_w = idim_w + pad_w
+ return [odim_h, odim_w]
def get_exp_cycles(self):
- odim = self.get_padded_odim()
+ odim_h, odim_w = self.get_padded_odim()
channels = self.get_nodeattr("NumChannels")
simd = self.get_nodeattr("SIMD")
batch_size = self.get_nodeattr("numInputVectors")
- exp_cycles = (channels / simd) * batch_size * odim * odim
+ exp_cycles = (channels / simd) * batch_size * odim_h * odim_w
return int(exp_cycles)
def get_normal_input_shape(self):
- idim = self.get_nodeattr("ImgDim")
+ idim_h, idim_w = self.get_nodeattr("ImgDim")
num_ch = self.get_nodeattr("NumChannels")
-
- ishape = (1, idim, idim, num_ch)
+ ishape = (1, idim_h, idim_w, num_ch)
return ishape
def get_normal_output_shape(self):
- odim = self.get_padded_odim()
+ odim_h, odim_w = self.get_padded_odim()
num_ch = self.get_nodeattr("NumChannels")
- oshape = (1, odim, odim, num_ch)
+ oshape = (1, odim_h, odim_w, num_ch)
return oshape
def get_folded_input_shape(self):
@@ -148,20 +158,53 @@ class FMPadding_Batch(HLSCustomOp):
self.code_gen_dict["$GLOBALS$"] = ['#include "streamtools.h"']
def defines(self, var):
- self.code_gen_dict["$DEFINES$"] = [
- """#define ImgDim1 {}\n#define OutputDim1 {}\n
- #define Padding1 {}\n#define NumChannels1 {}\n
- #define PaddingStyle1 {}\n#define numReps {}
- #define SIMD1 {}\n""".format(
- self.get_nodeattr("ImgDim"),
- self.get_padded_odim(),
- self.get_nodeattr("Padding"),
- self.get_nodeattr("NumChannels"),
- self.get_nodeattr("PaddingStyle"),
- self.get_nodeattr("numInputVectors"),
- self.get_nodeattr("SIMD"),
- )
- ]
+ idim_h, idim_w = self.get_nodeattr("ImgDim")
+ odim_h, odim_w = self.get_padded_odim()
+ pad = self.get_nodeattr("Padding")
+ pad_h = pad[0] + pad[2]
+ pad_w = pad[1] + pad[3]
+ is_square = idim_h == idim_w
+
+ if is_square:
+ assert (
+ pad_h == pad_w
+ ), "Only equal padding along the dimensions for square images is supported"
+ self.code_gen_dict["$DEFINES$"] = [
+ """#define ImgDim1 {}\n#define OutputDim1 {}\n
+ #define Padding1 {}\n#define NumChannels1 {}\n
+ #define SIMD1 {}\n#define PaddingStyle1 {}\n
+ #define numReps {}\n""".format(
+ idim_h,
+ odim_h,
+ pad_h,
+ self.get_nodeattr("NumChannels"),
+ self.get_nodeattr("SIMD"),
+ self.get_nodeattr("PaddingStyle"),
+ self.get_nodeattr("numInputVectors"),
+ )
+ ]
+ else:
+ self.code_gen_dict["$DEFINES$"] = [
+ """
+ #define OutputDim1_x {}\n
+ #define OutputDim1_y {}\n
+ #define Padding1_x {}\n
+ #define Padding1_y {}\n
+ #define NumChannels1 {}\n
+ #define SIMD1 {}\n
+ #define PaddingStyle1 {}\n
+ #define numReps {}\n
+ """.format(
+ odim_w,
+ odim_h,
+ pad_w,
+ pad_h,
+ self.get_nodeattr("NumChannels"),
+ self.get_nodeattr("SIMD"),
+ self.get_nodeattr("PaddingStyle"),
+ self.get_nodeattr("numInputVectors"),
+ )
+ ]
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
@@ -193,12 +236,26 @@ class FMPadding_Batch(HLSCustomOp):
def docompute(self):
in_t = self.get_input_datatype().get_hls_datatype_str()
node = self.onnx_node
- self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,SIMD1,
- {}, PaddingStyle1> (in0, out, numReps);""".format(
- node.op_type, in_t
- )
- ]
+
+ idim_h, idim_w = self.get_nodeattr("ImgDim")
+ is_square = idim_h == idim_w
+
+ if is_square:
+ hls_call = node.op_type
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,SIMD1,
+ {}, PaddingStyle1> (in0, out, numReps);""".format(
+ hls_call, in_t
+ )
+ ]
+ else:
+ hls_call = "FMPadding_nonsquare_Batch"
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """{}<OutputDim1_x, OutputDim1_y, Padding1_x, Padding1_y, NumChannels1,
+ SIMD1, {}, PaddingStyle1> (in0, out, numReps);""".format(
+ hls_call, in_t
+ )
+ ]
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
@@ -270,7 +327,7 @@ class FMPadding_Batch(HLSCustomOp):
assert (
inp.shape == exp_ishape
), """Input shape doesn't
- match expected shape (1, ImgDim, ImgDim, NumChannels)."""
+ match expected shape (1, ImgDim_h, ImgDim_w, NumChannels)."""
export_idt = self.get_input_datatype()
reshaped_input = inp.reshape(folded_ishape)
@@ -316,4 +373,4 @@ class FMPadding_Batch(HLSCustomOp):
assert (
context[node.output[0]].shape == exp_oshape
), """Output shape doesn't match expected shape
- (1, OutputDim, OutputDim, NumChannels)."""
+ (1, OutputDim_H, OutputDim_W, NumChannels)."""
diff --git a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
index b2835d578b03ee689330d53a9a7b233c9b9f4222..ab47b300136bd95622f064b30e3bbaae76a61597 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
@@ -54,15 +54,20 @@ target_clk_ns = 10
def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_style):
+ pad_h = padding[0] + padding[2]
+ pad_w = padding[1] + padding[3]
+ idim_h, idim_w = idim
+
assert pad_style == 2, "only pad_style == 2 supported in hlslib"
- assert padding > 0, "Output dim should be greater than input dim"
- odim = idim + padding
+ assert pad_h > 0 or pad_w > 0, "Output dim should be greater than input dim"
+ odim_h = idim_h + pad_h
+ odim_w = idim_w + pad_w
inp = helper.make_tensor_value_info(
- "inp", TensorProto.FLOAT, [1, idim, idim, num_ch]
+ "inp", TensorProto.FLOAT, [1, idim_h, idim_w, num_ch]
)
outp = helper.make_tensor_value_info(
- "outp", TensorProto.FLOAT, [1, odim, odim, num_ch]
+ "outp", TensorProto.FLOAT, [1, odim_h, odim_w, num_ch]
)
FMPadding = helper.make_node(
@@ -94,9 +99,9 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_sty
# input image dimension
-@pytest.mark.parametrize("idim", [8])
+@pytest.mark.parametrize("idim", [[8, 8], [10, 8]])
# number of rows and number of cols to add
-@pytest.mark.parametrize("pad", [2, 3])
+@pytest.mark.parametrize("pad", [[1, 1, 1, 1], [1, 1, 2, 2], [1, 3, 2, 3]])
# number of channels
@pytest.mark.parametrize("num_ch", [2, 4])
# Input parallelism
@@ -112,10 +117,22 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_sty
def test_fpgadataflow_fmpadding(idim, pad, num_ch, simd, pad_style, idt, mode):
if num_ch % simd != 0:
pytest.skip(" num_ch % simd != 0, skipping")
+
+ idim_h, idim_w = idim
+ pad_h = pad[0] + pad[2]
+ pad_w = pad[1] + pad[3]
+
+ if idim_h == idim_w and pad_h != pad_w:
+ pytest.skip(
+ """Only equal padding along the dimensions for square images
+ is supported, skipping"""
+ )
+
# generate input data
- x = gen_finn_dt_tensor(idt, [1, idim, idim, num_ch])
+ x = gen_finn_dt_tensor(idt, [1, idim_h, idim_w, num_ch])
input_dict = {"inp": x}
- odim = idim + pad
+ odim_h = idim_h + pad_h
+ odim_w = idim_w + pad_w
model = make_single_fmpadding_modelwrapper(idim, pad, num_ch, simd, idt, pad_style)
model = model.transform(InferShapes())
@@ -129,24 +146,26 @@ def test_fpgadataflow_fmpadding(idim, pad, num_ch, simd, pad_style, idt, mode):
model = model.transform(HLSSynthIP())
model = model.transform(PrepareRTLSim())
y_produced = oxe.execute_onnx(model, input_dict)["outp"]
- expected_oshape = (1, odim, odim, num_ch)
+ expected_oshape = (1, odim_h, odim_w, num_ch)
assert y_produced.shape == expected_oshape
# calculate reference
# calculate correct pad according to parameters
if pad_style == 2:
- if pad % 2 == 0:
- pad_up = pad // 2
- pad_left = pad // 2
+ if pad_h % 2 == 0:
+ pad_up = pad_h // 2
+ else:
+ pad_up = pad_h // 2 + 1
+ if pad_w % 2 == 0:
+ pad_left = pad_w // 2
else:
- pad_up = pad // 2 + 1
- pad_left = pad // 2 + 1
+ pad_left = pad_w // 2 + 1
else:
- pad_up = pad // 2
- pad_left = pad // 2
+ pad_up = pad_h // 2
+ pad_left = pad_w // 2
- pad_down = pad - pad_up
- pad_right = pad - pad_left
+ pad_down = pad_h - pad_up
+ pad_right = pad_w - pad_left
y_expected = np.pad(
x, ((0, 0), (pad_up, pad_down), (pad_left, pad_right), (0, 0)), "constant"