From fa9ed7853441b9b2d40e48e6c5f799bfa71753e0 Mon Sep 17 00:00:00 2001
From: auphelia <jakobapk@web.de>
Date: Fri, 3 Apr 2020 18:19:24 +0100
Subject: [PATCH] [HLSCustomOp] Add dynamically setting of AP_INT_MAX_W
---
src/finn/custom_op/fpgadataflow/__init__.py | 7 +++++++
.../fpgadataflow/convolutioninputgenerator.py | 11 ++++++++---
.../custom_op/fpgadataflow/streamingfclayer_batch.py | 11 ++++++++---
.../custom_op/fpgadataflow/streamingmaxpool_batch.py | 3 +++
src/finn/custom_op/fpgadataflow/templates.py | 7 +++++--
5 files changed, 31 insertions(+), 8 deletions(-)
diff --git a/src/finn/custom_op/fpgadataflow/__init__.py b/src/finn/custom_op/fpgadataflow/__init__.py
index 10f8b7fee..32c09d655 100644
--- a/src/finn/custom_op/fpgadataflow/__init__.py
+++ b/src/finn/custom_op/fpgadataflow/__init__.py
@@ -99,6 +99,7 @@ class HLSCustomOp(CustomOp):
# generate top cpp file for ip generation
path = self.get_nodeattr("code_gen_dir_ipgen")
+ self.code_gen_dict["$AP_INT_MAX_W$"] = [str(self.get_ap_int_max_w())]
self.generate_params(model, path)
self.global_includes()
self.defines("ipgen")
@@ -156,6 +157,7 @@ class HLSCustomOp(CustomOp):
"""Generates c++ code for simulation (npysim)."""
node = self.onnx_node
path = self.get_nodeattr("code_gen_dir_npysim")
+ self.code_gen_dict["$AP_INT_MAX_W$"] = [str(self.get_ap_int_max_w())]
self.generate_params(model, path)
self.global_includes()
self.defines("npysim")
@@ -429,3 +431,8 @@ compilation transformations?
def get_outstream_width(self):
"""Returns output stream width, if implemented."""
raise Exception("get_outstream_width not implemented for this op")
+
+ def get_ap_int_max_w(self):
+ instream = self.get_instream_width()
+ outstream = self.get_outstream_width()
+ return max([instream, outstream])
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
index e2a2f90b8..fe48af67f 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
@@ -73,6 +73,10 @@ class ConvolutionInputGenerator(HLSCustomOp):
ishape = (1, ifm_dim, ifm_dim, ifm_ch)
return ishape
+ def get_folded_input_shape(self):
+ """Assumption: No folding on input"""
+ return self.get_normal_input_shape()
+
def get_normal_output_shape(self):
k = self.get_nodeattr("ConvKernelDim")
ifm_dim = self.get_nodeattr("IFMDim")
@@ -272,9 +276,10 @@ class ConvolutionInputGenerator(HLSCustomOp):
def defines(self, var):
numReps = 1
self.code_gen_dict["$DEFINES$"] = [
- """#define ConvKernelDim1 {}\n #define IFMChannels1 {}
- #define Input_precision1 {}\n #define IFMDim1 {}\n #define OFMDim1 {}
- #define SIMD1 {}\n #define Stride1 {}\n #define numReps {}""".format(
+ """#define ConvKernelDim1 {}\n #define IFMChannels1 {}\n
+ #define Input_precision1 {}\n #define IFMDim1 {}\n
+ #define OFMDim1 {}\n #define SIMD1 {}\n
+ #define Stride1 {}\n #define numReps {}""".format(
self.get_nodeattr("ConvKernelDim"),
self.get_nodeattr("IFMChannels"),
self.get_input_datatype().bitwidth(),
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index 00b8287a3..301a39973 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -270,6 +270,11 @@ class StreamingFCLayer_Batch(HLSCustomOp):
wp = self.get_weight_datatype().bitwidth()
return pe * simd * wp
+ def get_ap_int_max_w(self):
+ temp_value = super().get_ap_int_max_w()
+ weightstream = self.get_weightstream_width()
+ return max([weightstream, temp_value])
+
def get_folded_input_shape(self):
mw = self.get_nodeattr("MW")
simd = self.get_nodeattr("SIMD")
@@ -714,9 +719,9 @@ class StreamingFCLayer_Batch(HLSCustomOp):
numInputVectors = list(self.get_nodeattr("numInputVectors"))
numReps = np.prod(numInputVectors)
self.code_gen_dict["$DEFINES$"] = [
- """#define MW1 {}\n #define MH1 {}\n #define SIMD1 {}\n
- #define PE1 {}\n #define WMEM1 {}\n #define TMEM1 {}\n
- #define numReps {}""".format(
+ """#define MW1 {}\n #define MH1 {}\n
+ #define SIMD1 {}\n #define PE1 {}\n #define WMEM1 {}\n
+ #define TMEM1 {}\n #define numReps {}""".format(
self.get_nodeattr("MW"),
self.get_nodeattr("MH"),
self.get_nodeattr("SIMD"),
diff --git a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
index 804da50f5..d7378b46c 100644
--- a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
@@ -64,6 +64,9 @@ class StreamingMaxPool_Batch(HLSCustomOp):
ishape = (1, ifm_dim, ifm_dim, ifm_ch)
return ishape
+ def get_folded_input_shape(self):
+ return self.get_normal_input_shape()
+
def get_normal_output_shape(self):
k = self.get_nodeattr("PoolDim")
ifm_dim = self.get_nodeattr("ImgDim")
diff --git a/src/finn/custom_op/fpgadataflow/templates.py b/src/finn/custom_op/fpgadataflow/templates.py
index 90a54b019..44a4def41 100644
--- a/src/finn/custom_op/fpgadataflow/templates.py
+++ b/src/finn/custom_op/fpgadataflow/templates.py
@@ -29,11 +29,12 @@
# template for single node execution
docompute_template = """
-#define AP_INT_MAX_W 16384
+#define AP_INT_MAX_W $AP_INT_MAX_W$
#include "cnpy.h"
#include "npy2apintstream.hpp"
#include <vector>
#include "bnn-library.h"
+
// includes for network parameters
$GLOBALS$
@@ -60,8 +61,10 @@ $SAVEASCNPY$
# cpp file
ipgen_template = """
-#define AP_INT_MAX_W 4096
+#define AP_INT_MAX_W $AP_INT_MAX_W$
+
#include "bnn-library.h"
+
// includes for network parameters
$GLOBALS$
--
GitLab