fixed conv's unittest (wrong shape for bias), added bank opt python s…

…cripts from DP-FGPA-V1

docs/banks.md

@@ -0,0 +1,10 @@
+# Memory Bank Optimizer
+`scripts/bank_optimizer_v2.py` is created to find optimized memory banks for the kernels with the objective of minimizing the number of `DataMover` kernel launches needed to execute the computation graph of the selected `TopModelArch`.
+
+# Instructions
+1. Run the selected `CModel` on FPGA for `hostlog_0trace.log` to be created.
+2. Open `hostlog_0trace.log` and find the lines that after `Dumping bank-crossing logs for` and append them.
+3. Copy the these lines into `bank_optimizer_v2.py`, method `get_objective` and line `objective =`
+4. Assign the allowed banks per kernel like `banks_transpose=[1,2]` to allow banks one and two to be selected for kernel `Transpose`, or `banks_transpose=[1]` to force the kernel to use only the bank one.
+5. Run the script.
+6. Use the output to configure `config` submodule of the main `DeepPoint-V2-FPGA` repository and then rebuild the FPGA image.

docs/results.md

@@ -0,0 +1,6 @@
+# Results
+Here the results of the various FPGA runs on AWS F1 could be found.
+To avoid including the related files in the git repository, they are lined below as github release archives.
+
+## Archive Links
+///TODO

readme.md

@@ -8,6 +8,7 @@ This repository contains the second version of the code base for Xilinx SDAccel
 - [Debugging The Project](docs/debug.md)
 - [Deploying on AWS-F1](docs/aws.md)
 - [Tips & Tricks!](docs/tips.md)
+- [Results](docs/results.md)
 
 # Build System
 As easy as it is to use SDx GUI, it is recommended to use provided cmake scripts to run synthesis and build the binaries for both the selected FPGA platform and the host.
@@ -18,7 +19,7 @@ This project relies on these software/libraries(These should be installed on the
 Xilinx SDAccel 2019.1(Tested), 2018.3 2018.2 2017.4(Not Tested)
 Xilinx XRT
 python3(Symlinked as `python3`)
-CMake3 (>3.10, Do **not** use default CMake package available on AWS-F1)
+CMake3 (>3.10, Do **not** use the default CMake package available on AWS-F1)
 Bash (>4.0, Dash and others are not tested)
 devtoolset-7 (>7.0, For C++14 support)
 ```

scripts/kernel_obj.py

@@ -0,0 +1,66 @@
+import numpy as np
+
+
+class SlrStats:
+    def __init__(self, slr_index, util_bram, util_dsp, util_ff, util_lut, util_uram):
+        self.slr_index = slr_index
+        self.util_bram = util_bram
+        self.util_dsp = util_dsp
+        self.util_ff = util_ff
+        self.util_lut = util_lut
+        self.util_uram = util_uram
+
+    def __add__(self, other):
+        if not(self.slr_index == -1 and other.slr_index == -1):
+            assert self.slr_index == other.slr_index
+        return SlrStats(
+            self.slr_index,
+            self.util_bram + other.util_bram,
+            self.util_dsp  + other.util_dsp ,
+            self.util_ff   + other.util_ff  ,
+            self.util_lut  + other.util_lut ,
+            self.util_uram + other.util_uram
+        )
+
+    def __str__(self):
+        return ''.join([
+            "BRAM=",str(self.util_bram),'%, ',
+            "DSP=",str(self.util_dsp),'%, ',
+            "FF=",str(self.util_ff),'%, ',
+            "LUT=",str(self.util_lut),'%, ',
+            "URAM=",str(self.util_uram),'%, ',
+        ])
+
+    def __lt__(self, other):
+        return self.util_bram < other.util_bram and \
+               self.util_dsp < other.util_dsp and \
+               self.util_ff < other.util_ff and \
+               self.util_lut < other.util_lut and \
+               self.util_uram < other.util_uram
+
+
+class KernelObj:
+    def __init__(self, kernel_name, possible_banks, util_bram, util_dsp, util_ff, util_lut, util_uram, assigned_bank=-1):
+        self.kernel_name = kernel_name
+        self.possible_banks = possible_banks
+        self.assigned_bank = assigned_bank
+        self.util_stats = SlrStats(-1, util_bram, util_dsp, util_ff, util_lut, util_uram)
+
+    def get_slr(self):
+        if self.assigned_bank == 1:
+            return 2
+        if self.assigned_bank == 2:
+            return 1
+        assert False
+
+    def clone(self, assigned_bank=-1):
+        return KernelObj(
+            self.kernel_name,
+            self.possible_banks,
+            self.util_stats.util_bram,
+            self.util_stats.util_dsp,
+            self.util_stats.util_ff,
+            self.util_stats.util_lut,
+            self.util_stats.util_uram,
+            assigned_bank)
+

test/ocltests/CMakeLists.txt

@@ -25,12 +25,12 @@ set(TEST_SOURCES
         #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwtile/test_ckwtile.cpp
         #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwtranspose/test_ckwtranspose.cpp
         #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwgather/test_ckwgather.cpp
-        ${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwreduce/test_ckwreduce.cpp
-        ${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_layermean/test_layermean.cpp
-        ${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_layervariance/test_layervariance.cpp
+        #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwreduce/test_ckwreduce.cpp
+        #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_layermean/test_layermean.cpp
+        #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_layervariance/test_layervariance.cpp
         #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwpadunpad/test_ckwpadunpad.cpp
         #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwtopk/test_ckwtopk.cpp
-        #${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwconv/test_ckwconv.cpp
+        ${CMAKE_SOURCE_DIR}/test/ocltests/unittests/test_ckwconv/test_ckwconv.cpp
         )
 
 

test/ocltests/unittests/test_ckwconv/test_ckwconv.cpp

@@ -29,7 +29,7 @@ bool ConvTest1(
 
 TEST(test_ckwconv, mixed1) {
   std::vector<bool> results = {
-      ConvTest1<float>({1,256,1,6},{1,1,6,16},{128}),
+      ConvTest1<float>({1,256,1,6},{1,1,6,16},{16}),
   };
 
   for(auto r:results){