From a976e5bc8a9d360485f8a7e2077dc39a58368c73 Mon Sep 17 00:00:00 2001
From: Lucian Petrica <lucianp@xilinx.com>
Date: Fri, 23 Jul 2021 11:58:47 +0100
Subject: [PATCH] Removed template for streaming thresholds (moved to hlslib);
added h/w dimensions for imdim (non-square)
---
src/finn/custom_op/fpgadataflow/templates.py | 53 -------------------
.../fpgadataflow/thresholding_batch.py | 22 ++++----
2 files changed, 11 insertions(+), 64 deletions(-)
diff --git a/src/finn/custom_op/fpgadataflow/templates.py b/src/finn/custom_op/fpgadataflow/templates.py
index 40221ce3b..7f8dbc787 100644
--- a/src/finn/custom_op/fpgadataflow/templates.py
+++ b/src/finn/custom_op/fpgadataflow/templates.py
@@ -355,56 +355,3 @@ $LAYER_NAME$
endmodule
"""
-
-decoupled_thresholding_template = """
-template <
- unsigned ImgDim, unsigned NumChannels, unsigned PE,
- typename TSrcI = Identity, typename TDstI = Identity,
- int ActVal=0, typename TT, unsigned int NumSteps,
- typename TI, typename TO>
-void Thresholding_Stream_Batch(hls::stream<TI> &in,
- hls::stream<TO> &out,
- hls::stream<ap_uint<PE*NumSteps*TT::width>> &weight,
- int const reps)
-{
-
- // how many different rows each neuron will compute
- // alternatively: number of vertical matrix chunks
- unsigned const NF = NumChannels / PE;
-
- 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
- // of smaller nested loops) to get the pipelinening the way we want
- for (unsigned i = 0; i < reps * ImgDim * ImgDim * NF; i++)
- {
- #pragma HLS PIPELINE II=1
-
- ap_uint<PE*NumSteps*TT::width> packed_thr;
- packed_thr = weight.read();
- // slicer to get 1 PE's worth of thresholds
- auto const pe_slicer = Slice<ap_uint<NumSteps*TT::width>>()(packed_thr);
-
- TI inElem;
- inElem = in.read();
- auto outElem = TDstI().template operator()<TO>();
-
- for (unsigned pe = 0; pe < PE; pe++)
- {
-#pragma HLS UNROLL
- // slicer to get individual thresholds
- auto const thr_slicer = Slice<TT>()(pe_slicer(pe, 0));
- for (unsigned nt = 0; nt < NumSteps; nt++)
- {
- #pragma HLS UNROLL
- internal_thr.m_thresholds[pe][0][nt] = thr_slicer(nt, 0);
- }
-
- auto const act = TSrcI()(inElem);
- outElem(pe,0,1) = internal_thr.activate(0, pe, act(pe,0));
- }
- out.write(outElem);
- }
-}
-"""
diff --git a/src/finn/custom_op/fpgadataflow/thresholding_batch.py b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
index 2944aeaa3..b4dc32943 100644
--- a/src/finn/custom_op/fpgadataflow/thresholding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
@@ -621,10 +621,6 @@ class Thresholding_Batch(HLSCustomOp):
self.code_gen_dict["$DEFINES$"].append(
"#define NumSteps1 %d" % self.get_nodeattr("numSteps")
)
- # TODO remove once Thresholding_Stream_Batch is in hlslib:
- self.code_gen_dict["$DEFINES$"].append(
- templates.decoupled_thresholding_template
- )
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
@@ -679,28 +675,32 @@ class Thresholding_Batch(HLSCustomOp):
node = self.onnx_node
ishape = self.get_folded_input_shape()
if len(ishape) == 3:
- imgdim = 1
+ imgdimh = 1
+ imgdimw = 1
elif len(ishape) == 5:
- imgdim = ishape[1]
+ imgdimh = ishape[1]
+ imgdimw = ishape[2]
else:
- raise Exception("""Unexpeted input shape""")
+ raise Exception("""Unexpected input shape""")
mem_mode = self.get_nodeattr("mem_mode")
if mem_mode == "const":
self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<{}, NumChannels1, PE1, {}, {}>
+ """{}<{}, {}, NumChannels1, PE1, {}, {}>
(in0, out, threshs, numReps);""".format(
node.op_type,
- imgdim,
+ imgdimh,
+ imgdimw,
tmpl_args["TSrcI"],
tmpl_args["TDstI"],
)
]
elif mem_mode == "decoupled":
self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<{}, NumChannels1, PE1, {}, {}, ActVal1, ThresType1, NumSteps1>
+ """{}<{}, {}, NumChannels1, PE1, {}, {}, ActVal1, ThresType1, NumSteps1>
(in0, out, weights, numReps);""".format(
"Thresholding_Stream_Batch",
- imgdim,
+ imgdimh,
+ imgdimw,
tmpl_args["TSrcI"],
tmpl_args["TDstI"],
)
--
GitLab