From b5520dc15175bd91517b2dd3209cee9893fe63a0 Mon Sep 17 00:00:00 2001
From: Felix Jentzsch <fepaje@mail.upb.de>
Date: Tue, 25 May 2021 16:44:00 +0200
Subject: [PATCH] Further cleanup
---
docker/finn_entrypoint.sh | 2 +-
.../fpgadataflow/streamingmaxpool_batch.py | 117 +++++++-----------
.../fpgadataflow/convert_to_hls_layers.py | 39 ++----
.../test_layer_streaming_maxpool_batch.py | 50 ++++----
4 files changed, 81 insertions(+), 127 deletions(-)
diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 3933d8918..46a294cd7 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -16,7 +16,7 @@ FINN_BASE_COMMIT=ac0b86a63eb937b869bfa453a996a8a8b8506546
FINN_EXP_COMMIT=e9f97dcdb4db2f889b0f36af079a6a1792b7d4de
BREVITAS_COMMIT=14abbe1e7ef82485d79415871fcf5766b0a40a00
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=4d74baefa79df48b5a0348d63f39a26df075de51
+HLSLIB_COMMIT=b37337c571b98f40423020bc79f97e189f2661d5
PYVERILATOR_COMMIT=e2ff74030de3992dcac54bf1b6aad2915946e8cb
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
diff --git a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
index a15c91af1..dba279878 100644
--- a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
@@ -58,6 +58,23 @@ class StreamingMaxPool_Batch(HLSCustomOp):
"""Returns FINN DataType of output."""
return DataType[self.get_nodeattr("dataType")]
+ def get_1d_attrs_normalized(self):
+ # support both (1, D) and (D, 1) cases transparently:
+ # assume the dummy ('1') dimension is the Y-dimension, i.e.
+ # images and kernels (and their attributes) of dimension
+ # [H, W] = [Y, X] = [D, 1] or [1, D] are always mapped to [1, D]
+ ifm_dim = self.get_nodeattr("ImgDim")
+ k = self.get_nodeattr("PoolDim")
+ ifm_ch = self.get_nodeattr("NumChannels")
+ if ifm_dim[1] == 1:
+ ifm_dim = ifm_dim[::-1]
+ k = k[::-1]
+ return (ifm_dim, k, ifm_ch)
+
+ def is_1d(self):
+ ifm_dim, k, ifm_ch = self.get_1d_attrs_normalized()
+ return (ifm_dim[0] == 1) and (k[0] == 1)
+
def get_normal_input_shape(self):
ifm_dim_h, ifm_dim_w = self.get_nodeattr("ImgDim")
ifm_ch = self.get_nodeattr("NumChannels")
@@ -73,8 +90,8 @@ class StreamingMaxPool_Batch(HLSCustomOp):
return tuple(ret)
def get_normal_output_shape(self):
- k_h, k_w = self.get_nodeattr("PoolDim")
ifm_dim_h, ifm_dim_w = self.get_nodeattr("ImgDim")
+ k_h, k_w = tuple(self.get_nodeattr("PoolDim"))
ifm_ch = self.get_nodeattr("NumChannels")
stride_h = k_h
stride_w = k_w
@@ -100,13 +117,12 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def get_exp_cycles(self):
# derived from StreamingMaxPool_Batch loop nest
- k_h, k_w = self.get_nodeattr("PoolDim")
- ifm_dim_h, ifm_dim_w = self.get_nodeattr("ImgDim")
- # TODO: correct formula
- if k_h == k_w: # todo: better condition
- return int(ifm_dim_h * (ifm_dim_h + (ifm_dim_h / k_h)))
+ ifm_dim, k, ifm_ch = self.get_1d_attrs_normalized()
+ if self.is_1d():
+ return int(ifm_dim[1] + k[1])
else:
- return int((ifm_dim_h / k_h) * (k_h + 1))
+ # TODO: adjust inaccurate formula
+ return int(ifm_dim[1] * (ifm_dim[1] + (ifm_dim[1] / k[1])))
def get_instream_width(self):
dt_bits = self.get_input_datatype().bitwidth()
@@ -173,51 +189,17 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def defines(self, var):
numReps = 2
- k = self.get_nodeattr("PoolDim")
- ifm_dim = self.get_nodeattr("ImgDim")
-
- if k[0] == k[1]: # todo: better condition
- self.code_gen_dict["$DEFINES$"] = [
- """#define ImgDim {}\n #define PoolDim {}\n
- #define NumChannels {}\n #define numReps {}""".format(
- ifm_dim[0],
- k[0],
- self.get_nodeattr("NumChannels"),
- numReps,
- )
- ]
- else:
- # TODO: use the same convention als convinpgen?:
-
- # For the kernel, presenting the input data of size D as
- # [H, W] = [Y, X] = [1, D] or [D, 1]
- # effectively gives the same result. Because the
- # ConvolutionInputGenerator_NonSquare_Dilated(_dws) kernel currently only
- # supports dilation>1 along the X-axis and the
- # ConvolutionInputGenerator_NonSquare only works for stride>1 along the
- # X-axis, we are working with the following assumption:
- # the dummy ('1') dimension is the Y-dimension, i.e.
- # images and kernels (and their attributes) of dimension
- # [H, W] = [Y, X] = [D, 1] or [1, D] are always mapped to [1, D]
- if ifm_dim[1] == 1:
- ifm_dim = ifm_dim[::-1]
- k = k[::-1]
-
- ifm_dim_y, ifm_dim_x = ifm_dim
- k_y, k_x = k
-
- self.code_gen_dict["$DEFINES$"] = [
- """#define ImgDim_x {}\n #define ImgDim_y {}\n
- #define PoolDim_x {}\n #define PoolDim_y {}\n
- #define NumChannels {}\n #define numReps {}""".format(
- ifm_dim_x,
- ifm_dim_y,
- k_x,
- k_y,
- self.get_nodeattr("NumChannels"),
- numReps,
- )
- ]
+ ifm_dim, k, ifm_ch = self.get_1d_attrs_normalized()
+
+ self.code_gen_dict["$DEFINES$"] = [
+ """#define ImgDim {}\n #define PoolDim {}\n
+ #define NumChannels {}\n #define numReps {}""".format(
+ ifm_dim[1],
+ k[1],
+ self.get_nodeattr("NumChannels"),
+ numReps,
+ )
+ ]
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
@@ -248,33 +230,24 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def docompute(self):
dtype = self.get_input_datatype()
-
- k = self.get_nodeattr("PoolDim")
- # ifm_dim = self.get_nodeattr("ImgDim")
- if k[0] == k[1]: # todo: better condition
- if dtype.bitwidth() == 1:
+ if dtype.bitwidth() == 1:
+ if self.is_1d():
+ raise Exception("Binary 1d MaxPool not implemented on HLS backend")
+ else:
op = "StreamingMaxPool_Batch"
- self.code_gen_dict["$DOCOMPUTE$"] = [
- "%s<ImgDim, PoolDim, NumChannels>(in0, out, numReps);" % (op)
- ]
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "%s<ImgDim, PoolDim, NumChannels>(in0, out, numReps);" % (op)
+ ]
+ else:
+ if self.is_1d():
+ op = "StreamingMaxPool_Precision_Batch_1d"
else:
op = "StreamingMaxPool_Precision_Batch"
- dtype = self.get_input_datatype()
- dtype_hls = dtype.get_hls_datatype_str()
- minval_str = str(int(dtype.min()))
- self.code_gen_dict["$DOCOMPUTE$"] = [
- "%s<ImgDim, PoolDim, NumChannels, %s, %s>(in0, out, numReps);"
- % (op, dtype_hls, minval_str)
- ]
- else:
- # todo: add binary op
- op = "StreamingMaxPool_Precision_Batch_NonSquare"
dtype = self.get_input_datatype()
dtype_hls = dtype.get_hls_datatype_str()
minval_str = str(int(dtype.min()))
self.code_gen_dict["$DOCOMPUTE$"] = [
- """%s<ImgDim_x, ImgDim_y, PoolDim_x, PoolDim_y,
- NumChannels, %s, %s>(in0, out, numReps);"""
+ "%s<ImgDim, PoolDim, NumChannels, %s, %s>(in0, out, numReps);"
% (op, dtype_hls, minval_str)
]
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index d3989343f..e1c3eced7 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -235,40 +235,23 @@ class InferStreamingMaxPool(Transformation):
# mp_out_shape = model.get_tensor_shape(mp_output)
dt = model.get_tensor_datatype(mp_input)
mp_inst = getCustomOp(n)
- # stride = mp_inst.get_nodeattr("strides")[0]
k_h, k_w = mp_inst.get_nodeattr("kernel_shape")
- # pad = mp_inst.get_nodeattr("pads")[0]
ifm_ch = mp_in_shape[-1]
ifm_dim_h = mp_in_shape[1]
ifm_dim_w = mp_in_shape[2]
- # ofm_dim = mp_out_shape[1]
if ifm_dim_h % k_h == 0 and ifm_dim_w % k_w == 0:
# create equivalent StreamingMaxPool_Batch node
- # TODO support non-k strides
- if k_h == k_w: # todo: better condition or none at all
- new_node = helper.make_node(
- "StreamingMaxPool_Batch",
- [mp_input],
- [mp_output],
- domain="finn.custom_op.fpgadataflow",
- backend="fpgadataflow",
- PoolDim=k_h,
- NumChannels=ifm_ch,
- ImgDim=ifm_dim_h,
- dataType=dt.name,
- )
- else:
- new_node = helper.make_node(
- "StreamingMaxPool_Batch",
- [mp_input],
- [mp_output],
- domain="finn.custom_op.fpgadataflow",
- backend="fpgadataflow",
- PoolDim=(k_h, k_w),
- NumChannels=ifm_ch,
- ImgDim=(ifm_dim_h, ifm_dim_w),
- dataType=dt.name,
- )
+ new_node = helper.make_node(
+ "StreamingMaxPool_Batch",
+ [mp_input],
+ [mp_output],
+ domain="finn.custom_op.fpgadataflow",
+ backend="fpgadataflow",
+ PoolDim=(k_h, k_w),
+ NumChannels=ifm_ch,
+ ImgDim=(ifm_dim_h, ifm_dim_w),
+ dataType=dt.name,
+ )
graph.node.insert(node_ind, new_node)
# remove old nodes
graph.node.remove(n)
diff --git a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py b/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
index 790c0cb7e..2220bd29e 100644
--- a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
+++ b/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
@@ -43,6 +43,7 @@ from finn.transformation.general import GiveUniqueNodeNames
from finn.util.basic import gen_finn_dt_tensor
from finn.analysis.fpgadataflow.exp_cycles_per_layer import exp_cycles_per_layer
from finn.custom_op.registry import getCustomOp
+import numpy as np
def make_single_maxpoolnhwc_modelwrapper(k, ifm_ch, ifm_dim, ofm_dim, idt):
@@ -120,37 +121,38 @@ def prepare_inputs(input_tensor):
# input datatype
-# @pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT2])
-@pytest.mark.parametrize("idt", [DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT4])
+# 1d maxpool
+@pytest.mark.parametrize("dim_1d", [False, True])
# kernel size
-@pytest.mark.parametrize(
- "k",
- [
- (2, 1),
- ],
-) # (4,4)])
+@pytest.mark.parametrize("k", [2, 4])
# input dimension
-@pytest.mark.parametrize(
- "ifm_dim",
- [
- (1024, 1),
- ],
-) # (6,6), (8,8)])
+@pytest.mark.parametrize("ifm_dim", [4, 8])
# input channels
-@pytest.mark.parametrize("ifm_ch", [1, 3])
+@pytest.mark.parametrize("ifm_ch", [1, 3]) # 1,3
# execution mode
-# @pytest.mark.parametrize("exec_mode", ["rtlsim", "cppsim"])
-@pytest.mark.parametrize("exec_mode", ["rtlsim"])
+@pytest.mark.parametrize("exec_mode", ["rtlsim", "cppsim"])
@pytest.mark.slow
@pytest.mark.vivado
-def test_fpgadataflow_streamingmaxpool(idt, k, ifm_dim, ifm_ch, exec_mode):
- k_h, k_w = k
- ifm_dim_h, ifm_dim_w = ifm_dim
+def test_fpgadataflow_streamingmaxpool(idt, dim_1d, k, ifm_dim, ifm_ch, exec_mode):
+ ifm_dim_h = ifm_dim
+ k_h = k
+ if dim_1d:
+ ifm_dim_w = 1
+ k_w = 1
+ else:
+ ifm_dim_w = ifm_dim_h
+ k_w = k_h
+ ifm_dim = (ifm_dim_h, ifm_dim_w)
+ k = (k_h, k_w)
+
stride_h = k_h
stride_w = k_w
ofm_dim_h = int(((ifm_dim_h - k_h) / stride_h) + 1)
ofm_dim_w = int(((ifm_dim_w - k_w) / stride_w) + 1)
ofm_dim = (ofm_dim_h, ofm_dim_w)
+ if idt == DataType.BIPOLAR and dim_1d:
+ pytest.skip("Skipping binary StreamingMaxPool_1d (not implemented)")
if ifm_dim_h % k_h != 0 or ifm_dim_w % k_w != 0:
pytest.skip("Skipping StreamingMaxPool test w/ ImgDim % PoolDim != 0")
@@ -186,9 +188,5 @@ def test_fpgadataflow_streamingmaxpool(idt, k, ifm_dim, ifm_ch, exec_mode):
cycles_rtlsim = inst.get_nodeattr("cycles_rtlsim")
exp_cycles_dict = model.analysis(exp_cycles_per_layer)
exp_cycles = exp_cycles_dict[node.name]
- # DEBUG:
- print("expected vs rtlsim cycles")
- print(exp_cycles)
- print(cycles_rtlsim)
- # assert np.isclose(exp_cycles, cycles_rtlsim, atol=15)
- # assert exp_cycles != 0
+ assert np.isclose(exp_cycles, cycles_rtlsim, atol=15)
+ assert exp_cycles != 0
--
GitLab