From 8497814bf2834060cef05bf6832497e022b2e881 Mon Sep 17 00:00:00 2001
From: Yaman Umuroglu <yamanu@xilinx.com>
Date: Sun, 17 Jan 2021 01:51:00 +0100
Subject: [PATCH] Switch hlslib comparison functions (#263)
* [Test] add lfc to end2end tests
* [Deps] update hlslib to get comp:: fxns
* [HLS] use comp:: comparators instead of std::
* [VVAU] hlslib now uses inner prod dim instead of K
* [Thres] manually workaround vivado_hls bug for T[0][0]=0
---
docker/finn_entrypoint.sh | 2 +-
.../fpgadataflow/channelwise_op_batch.py | 2 +-
.../fpgadataflow/streamingfclayer_batch.py | 9 ++++++-
src/finn/custom_op/fpgadataflow/templates.py | 2 +-
.../fpgadataflow/thresholding_batch.py | 9 ++++++-
.../vector_vector_activate_batch.py | 17 +++++++++---
src/finn/util/test.py | 1 +
tests/end2end/test_end2end_bnn_pynq.py | 27 ++++++++++++++++++-
.../test_fpgadataflow_thresholding.py | 4 +++
9 files changed, 63 insertions(+), 10 deletions(-)
diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index ee8d88789..1ed8875e8 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -15,7 +15,7 @@ gecho () {
FINN_BASE_COMMIT=1363981654009067790d5f2d0c3dd303b5fa05cb
BREVITAS_COMMIT=aff49758ec445d77c75721c7de3091a2a1797ca8
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=cfafe11a93b79ab1af7529d68f08886913a6466e
+HLSLIB_COMMIT=2e49322d1bbc4969ca293843bda1f3f9c05456fc
PYVERILATOR_COMMIT=e2ff74030de3992dcac54bf1b6aad2915946e8cb
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
diff --git a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
index 635f37d56..097ec336f 100644
--- a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
+++ b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
@@ -363,7 +363,7 @@ class ChannelwiseOp_Batch(HLSCustomOp):
# get desired function
func = self.get_nodeattr("Func")
if func == "cmp_le":
- func_str = "std::less_equal"
+ func_str = "comp::less_equal"
elif func == "cmp_ge":
func_str = "std::greater_equal"
elif func == "add":
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index 118d66822..8868002c9 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -643,6 +643,13 @@ class StreamingFCLayer_Batch(HLSCustomOp):
# ensure all thresholds are integer
assert (orig_thres_matrix.astype(np.int32) == orig_thres_matrix).all()
ret = orig_thres_matrix
+ # workaround for vivado_hls threshold bug
+ if ret[0][0] == 0:
+ ret = np.copy(ret)
+ ret[0][0] = 1
+ warnings.warn(
+ "Setting 0-valued first threshold to 1 to avoid vivado_hls bug"
+ )
# ensure channels = mh , duplicating if necessary
if ret.shape[0] == 1:
ret = np.tile(ret, (mh, 1))
@@ -846,7 +853,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
tdt_hls,
odt_hls,
self.get_nodeattr("ActVal"),
- "std::less_equal<%s>" % tdt_hls,
+ "comp::less_equal<%s>" % tdt_hls,
)
)
f_thresh.write(thresholds_hls_code)
diff --git a/src/finn/custom_op/fpgadataflow/templates.py b/src/finn/custom_op/fpgadataflow/templates.py
index 48ba43419..40221ce3b 100644
--- a/src/finn/custom_op/fpgadataflow/templates.py
+++ b/src/finn/custom_op/fpgadataflow/templates.py
@@ -372,7 +372,7 @@ void Thresholding_Stream_Batch(hls::stream<TI> &in,
// alternatively: number of vertical matrix chunks
unsigned const NF = NumChannels / PE;
- ThresholdsActivation<1, PE, NumSteps, TT, TO, ActVal, std::less_equal<TT>> internal_thr;
+ ThresholdsActivation<1, PE, NumSteps, TT, TO, ActVal, comp::less_equal<TT>> internal_thr;
#pragma HLS ARRAY_PARTITION variable=internal_thr.m_thresholds complete dim=0
// everything merged into a common iteration space (one "big" loop instead
diff --git a/src/finn/custom_op/fpgadataflow/thresholding_batch.py b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
index ede3a2e58..30374a7d9 100644
--- a/src/finn/custom_op/fpgadataflow/thresholding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
@@ -334,6 +334,13 @@ class Thresholding_Batch(HLSCustomOp):
np.mod(orig_thres_matrix, 1), 0
).all(), "Need int threshold tensor"
ret = orig_thres_matrix
+ # workaround for vivado_hls threshold bug
+ if ret[0][0] == 0:
+ ret = np.copy(ret)
+ ret[0][0] = 1
+ warnings.warn(
+ "Setting 0-valued first threshold to 1 to avoid vivado_hls bug"
+ )
# ensure channels = mh , duplicating if necessary
if ret.shape[0] == 1:
ret = np.tile(ret, (mh, 1))
@@ -394,7 +401,7 @@ class Thresholding_Batch(HLSCustomOp):
tdt_hls,
odt_hls,
self.get_nodeattr("ActVal"),
- "std::less_equal<%s>" % tdt_hls,
+ "comp::less_equal<%s>" % tdt_hls,
)
)
f_thresh.write(thresholds_hls_code)
diff --git a/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py b/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
index c7e6466bc..9a897d9fa 100644
--- a/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
+++ b/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
@@ -278,6 +278,13 @@ class Vector_Vector_Activate_Batch(HLSCustomOp):
not as expected (2)."""
n_thres_steps = orig_thres_matrix.shape[1]
ret = orig_thres_matrix
+ # workaround for vivado_hls threshold bug
+ if ret[0][0] == 0:
+ ret = np.copy(ret)
+ ret[0][0] = 1
+ warnings.warn(
+ "Setting 0-valued first threshold to 1 to avoid vivado_hls bug"
+ )
# distribute rows between PEs
ret = interleave_matrix_outer_dim_from_partitions(ret, pe)
assert (
@@ -352,7 +359,7 @@ class Vector_Vector_Activate_Batch(HLSCustomOp):
tdt_hls,
odt_hls,
self.get_nodeattr("ActVal"),
- "std::less_equal<%s>" % tdt_hls,
+ "comp::less_equal<%s>" % tdt_hls,
)
)
f_thresh.write(thresholds_hls_code)
@@ -450,11 +457,13 @@ class Vector_Vector_Activate_Batch(HLSCustomOp):
def defines(self, var):
dim = self.get_nodeattr("Dim")
numReps = 1 * dim * dim
+ kernel = self.get_nodeattr("Kernel")
+ innerProdDim = kernel * kernel
self.code_gen_dict["$DEFINES$"] = [
- """#define Channels1 {}\n #define Kernel1 {}\n
+ """#define Channels1 {}\n #define InnerProdDim {}\n
#define SIMD1 1\n #define PE1 {}\n #define numReps {}""".format(
self.get_nodeattr("Channels"),
- self.get_nodeattr("Kernel"),
+ innerProdDim,
self.get_nodeattr("PE"),
numReps,
)
@@ -499,7 +508,7 @@ class Vector_Vector_Activate_Batch(HLSCustomOp):
threshs = "threshs"
node = self.onnx_node
self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<Channels1, Kernel1, SIMD1, PE1, 1, {}, {}, {}>
+ """{}<Channels1, InnerProdDim, SIMD1, PE1, 1, {}, {}, {}>
(in0, out, weights, {}, numReps, {});""".format(
node.op_type,
tmpl_args["TSrcI"],
diff --git a/src/finn/util/test.py b/src/finn/util/test.py
index 498b8b551..0a3475178 100644
--- a/src/finn/util/test.py
+++ b/src/finn/util/test.py
@@ -159,6 +159,7 @@ def get_trained_network_and_ishape(topology, wbits, abits):
topology_to_ishape = {
"tfc": (1, 1, 28, 28),
+ "lfc": (1, 1, 28, 28),
"cnv": (1, 3, 32, 32),
}
ishape = topology_to_ishape[topology]
diff --git a/tests/end2end/test_end2end_bnn_pynq.py b/tests/end2end/test_end2end_bnn_pynq.py
index 8ed544e8d..5f54eeacf 100644
--- a/tests/end2end/test_end2end_bnn_pynq.py
+++ b/tests/end2end/test_end2end_bnn_pynq.py
@@ -150,6 +150,27 @@ def fold_tfc(model):
return model
+def fold_lfc(model):
+ fc_layers = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
+ # (PE, SIMD, ramstyle) for each layer
+ config = [
+ (32, 49, "block"),
+ (64, 32, "auto"),
+ (32, 64, "auto"),
+ (10, 8, "distributed"),
+ ]
+ for fcl, (pe, simd, ramstyle) in zip(fc_layers, config):
+ fcl_inst = getCustomOp(fcl)
+ fcl_inst.set_nodeattr("PE", pe)
+ fcl_inst.set_nodeattr("SIMD", simd)
+ fcl_inst.set_nodeattr("ram_style", ramstyle)
+ # set parallelism for input quantizer to be same as first layer's SIMD
+ inp_qnt_node = model.get_nodes_by_op_type("Thresholding_Batch")[0]
+ inp_qnt = getCustomOp(inp_qnt_node)
+ inp_qnt.set_nodeattr("PE", 49)
+ return model
+
+
def fold_cnv_large(model):
fc_layers = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
# each tuple is (PE, SIMD) for a layer
@@ -208,6 +229,8 @@ def fold_cnv_small(model):
def get_folding_function(topology, wbits, abits):
if "tfc" in topology:
return fold_tfc
+ elif "lfc" in topology:
+ return fold_lfc
elif "cnv" in topology:
if wbits == 1 and abits == 1:
return fold_cnv_large
@@ -284,11 +307,13 @@ def topology2dataset(topology):
@pytest.mark.parametrize("wbits", [1, 2])
@pytest.mark.parametrize("abits", [1, 2])
-@pytest.mark.parametrize("topology", ["tfc", "cnv"])
+@pytest.mark.parametrize("topology", ["lfc", "tfc", "cnv"])
class TestEnd2End:
def test_export(self, topology, wbits, abits):
if wbits > abits:
pytest.skip("No wbits > abits end2end network configs for now")
+ if topology == "lfc" and wbits > 1:
+ pytest.skip("Skipping non-existing lfc configs")
(model, ishape) = get_trained_network_and_ishape(topology, wbits, abits)
chkpt_name = get_checkpoint_name(topology, wbits, abits, "export")
bo.export_finn_onnx(model, ishape, chkpt_name)
diff --git a/tests/fpgadataflow/test_fpgadataflow_thresholding.py b/tests/fpgadataflow/test_fpgadataflow_thresholding.py
index 5d46f4c3d..bbc7e8227 100644
--- a/tests/fpgadataflow/test_fpgadataflow_thresholding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_thresholding.py
@@ -122,6 +122,10 @@ def test_fpgadataflow_thresholding(idt, act, nf, ich, exec_mode, mem_mode):
odt = act
n_steps = act.get_num_possible_values() - 1
T = np.random.randint(idt.min(), idt.max() + 1, (ich, n_steps)).astype(np.float32)
+ # make the vivado_hls threshold bug appear (incorrect rtlsim result when first
+ # threshold of first channel is zero, while using BIPOLAR output)
+ if act == DataType.BIPOLAR:
+ T[0][0] = 0
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
--
GitLab