From 14929f41e948881a0ce20040c25ff2957928c7a0 Mon Sep 17 00:00:00 2001
From: auphelia <jakobapk@web.de>
Date: Wed, 3 Jan 2024 15:43:58 +0000
Subject: [PATCH] [Deconv] Update test and add comments to transformation

---
 .../infer_pixel_padding_deconv.py             | 25 ++++++++++++-------
 .../fpgadataflow/test_fpgadataflow_deconv.py  | 22 +++++++++++-----
 2 files changed, 32 insertions(+), 15 deletions(-)

diff --git a/src/finn/transformation/fpgadataflow/infer_pixel_padding_deconv.py b/src/finn/transformation/fpgadataflow/infer_pixel_padding_deconv.py
index 8642f5e0ef..8dbf7071fc 100644
--- a/src/finn/transformation/fpgadataflow/infer_pixel_padding_deconv.py
+++ b/src/finn/transformation/fpgadataflow/infer_pixel_padding_deconv.py
@@ -7,8 +7,14 @@
 
 
 class InferPixelPaddingDeconv(Transformation):
-    def __init__(self):
-        super().__init__()
+    """
+    Lowering and conversion of ConvTranspose (NCHW) nodes to
+    FMPadding_Pixel + Im2Col + MatMul (NHWC) surrounded by Transpose nodes
+    note: this transformation produces a mix of hw layers and non hw layers
+    to implement this on an FPGA the Im2Col and MatMul nodes need to be converted to hw layers
+    after applying this transformation and the resulting transpose nodes need to be streamlined.
+    See deconv test case under tests/fpgadataflow for an example.
+    """
 
     def apply(self, model):
         graph = model.graph
@@ -17,6 +23,14 @@ def apply(self, model):
         for n in graph.node:
             node_ind += 1
             if n.op_type == "ConvTranspose":
+                # conversion currently only supported for group=1
+                group = get_by_name(n.attribute, "group").i
+                if group != 1:
+                    warnings.warn(
+                        "%s : Only group=1 is currently supported. Can't infer PixelPaddingDeconv."
+                        % n.name
+                    )
+                    continue
                 deconv_input = n.input[0]
                 deconv_output = n.output[0]
                 idt = model.get_tensor_datatype(deconv_input)
@@ -25,13 +39,6 @@ def apply(self, model):
                 k_w = get_by_name(n.attribute, "kernel_shape").ints[1]
                 stride_h = get_by_name(n.attribute, "strides").ints[0]
                 stride_w = get_by_name(n.attribute, "strides").ints[1]
-                group = get_by_name(n.attribute, "group").i
-                if group != 1:
-                    warnings.warn(
-                        "%s : Only group=1 is currently supported. Can't infer PixelPaddingDeconv."
-                        % n.name
-                    )
-                    continue
                 weight_name = n.input[1]
                 W_conv = model.get_initializer(weight_name)
                 ifm_ch = model.get_tensor_shape(n.input[0])[1]  # assume NCHW
diff --git a/tests/fpgadataflow/test_fpgadataflow_deconv.py b/tests/fpgadataflow/test_fpgadataflow_deconv.py
index d5edec9b35..6c25be0f85 100644
--- a/tests/fpgadataflow/test_fpgadataflow_deconv.py
+++ b/tests/fpgadataflow/test_fpgadataflow_deconv.py
@@ -28,6 +28,7 @@
 
 import pytest
 
+import numpy as np
 import os
 from onnx import TensorProto, helper
 from qonnx.core.datatype import DataType
@@ -38,6 +39,7 @@
 from qonnx.util.basic import gen_finn_dt_tensor, qonnx_make_model
 
 import finn.core.onnx_exec as oxe
+from finn.analysis.fpgadataflow.exp_cycles_per_layer import exp_cycles_per_layer
 from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
 from finn.transformation.fpgadataflow.convert_to_hls_layers import (
     InferConvInpGen,
@@ -177,14 +179,12 @@ def test_fpgadataflow_deconv(idim, stride, ifm_ch, ofm_ch, simd, pe, k, padding,
 
     expected_oshape = (1, ofm_ch, odim_h, odim_w)
     y_expected = oxe.execute_onnx(ref_model, input_dict)["outp"]
+
     # cppsim
     if exec_mode == "cppsim":
         model = model.transform(PrepareCppSim())
         model = model.transform(CompileCppSim())
         model = model.transform(SetExecMode("cppsim"))
-        y_produced = oxe.execute_onnx(model, input_dict)["outp"]
-        assert y_produced.shape == expected_oshape
-        assert (y_produced == y_expected).all()
 
     # rtlsim
     else:
@@ -192,6 +192,16 @@ def test_fpgadataflow_deconv(idim, stride, ifm_ch, ofm_ch, simd, pe, k, padding,
         model = model.transform(HLSSynthIP())
         model = model.transform(PrepareRTLSim())
         model = model.transform(SetExecMode("rtlsim"))
-        y_produced = oxe.execute_onnx(model, input_dict)["outp"]
-        assert y_produced.shape == expected_oshape
-        assert (y_produced == y_expected).all()
+
+    y_produced = oxe.execute_onnx(model, input_dict)["outp"]
+    assert y_produced.shape == expected_oshape
+    assert (y_produced == y_expected).all()
+
+    if exec_mode == "rtlsim":
+        node = model.get_nodes_by_op_type("FMPadding_Pixel")[0]
+        inst = getCustomOp(node)
+        cycles_rtlsim = inst.get_nodeattr("cycles_rtlsim")
+        exp_cycles_dict = model.analysis(exp_cycles_per_layer)
+        exp_cycles = exp_cycles_dict[node.name]
+        assert np.isclose(exp_cycles, cycles_rtlsim, atol=10)
+        assert exp_cycles != 0