From c63316378101df5fb38e2d3f94a8545f1ff50242 Mon Sep 17 00:00:00 2001
From: Pete Warden <pete@petewarden.com>
Date: Sun, 10 May 2015 19:52:43 +0000
Subject: [PATCH] Fixes for Raspberry Pi 2

---
 source/Makefile                     | 11 +++++++++--
 source/src/lib/graph/buffer.cpp     | 13 ++++++++++++-
 source/src/lib/math/matrix_gemm.cpp |  8 ++++++++
 source/src/lib/pi/qpu_gemm.cpp      |  5 ++++-
 4 files changed, 33 insertions(+), 4 deletions(-)

diff --git a/source/Makefile b/source/Makefile
index 8525bff..afa9168 100644
--- a/source/Makefile
+++ b/source/Makefile
@@ -5,7 +5,7 @@ CPPFLAGS=
 
 .PRECIOUS: %.cdat
 
-LIBCPPFLAGS=-Ofast -I ./src/lib/include -I ./src/lib/graph -I ./src/lib/math -I ./src/lib/third_party -I ./src/lib/utility -I ./src/lib/svm -I ./src/lib/opengl -I ./src/lib -I ./src/include
+LIBCPPFLAGS=-Ofast -I ./src/lib/include -I ./src/lib/graph -I ./src/lib/math -I ./src/lib/third_party -I ./src/lib/utility -I ./src/lib/svm -I ./src/lib/opengl -I ./src/lib -I ./src/include -g
 LIBLDFLAG=
 LIBLDLIBS=
 
@@ -42,6 +42,13 @@ LIBCPPFLAGS += \
 -I./src/lib/pi
 endif
 
+ifeq ($(TARGET),pi2)
+LIBCPPFLAGS += \
+-mcpu=cortex-a7 \
+-mfpu=neon-vfpv4 \
+-mfloat-abi=hard
+endif
+
 ifeq ($(GEMM),pigl)
 LIBCPPFLAGS += \
 -I/usr/include \
@@ -63,7 +70,7 @@ ASMOBJS := $(subst .asm,.do,$(ASMSRCS))
 LIBOBJS += $(ASMOBJS)
 endif
 
-TOOLCPPFLAGS := -O3 -I ./src/include
+TOOLCPPFLAGS := -O3 -I ./src/include -g
 
 TOOLSRCS := $(shell find src/tool -name '*.cpp' -not -name '._*')
 TOOLOBJS := $(subst .cpp,.o,$(TOOLSRCS))
diff --git a/source/src/lib/graph/buffer.cpp b/source/src/lib/graph/buffer.cpp
index 3de05ec..a3e8708 100644
--- a/source/src/lib/graph/buffer.cpp
+++ b/source/src/lib/graph/buffer.cpp
@@ -31,6 +31,10 @@
 #include "mailbox.h"
 #endif // TARGET_PI
 
+#ifdef USE_EIGEN_GEMM
+#include <Eigen/Core>
+#endif // USE_EIGEN_GEMM
+
 static void buffer_do_save_to_image_file(Buffer* buffer, const char* filename);
 
 Buffer::Buffer(const Dimensions& dims) :
@@ -51,6 +55,8 @@ Buffer::Buffer(const Dimensions& dims) :
   }
   _gpuMemoryBase = mem_lock(_gpuMemoryHandle);
   _data = (jpfloat_t*)(mapmem(_gpuMemoryBase + GPU_MEM_MAP, byteCount));
+#elif defined(USE_EIGEN_GEMM)
+  _data = (jpfloat_t*)(Eigen::internal::aligned_malloc(byteCount));
 #else // TARGET_PI
   _data = (jpfloat_t*)(malloc(byteCount * 1));
 #endif // TARGET_PI
@@ -114,8 +120,10 @@ Buffer::Buffer(const Dimensions& dims, jpfloat_t min, jpfloat_t max, int bitsPer
   }
   _gpuMemoryBase = mem_lock(_gpuMemoryHandle);
   _quantizedData = (char*)(mapmem(_gpuMemoryBase + GPU_MEM_MAP, byteCount));
+#elif defined(USE_EIGEN_GEMM)
+  _quantizedData = (void*)(Eigen::internal::aligned_malloc(byteCount));
 #else // TARGET_PI
-  _quantizedData = (void*)(malloc(byteCount * 1));
+  _quantizedData = (void*)(malloc(byteCount));
 #endif // TARGET_PI
   _doesOwnData = true;
   setName("None");
@@ -137,6 +145,9 @@ Buffer::~Buffer()
     }
     mem_unlock(_gpuMemoryHandle);
     mem_free(_gpuMemoryHandle);
+#elif defined(USE_EIGEN_GEMM)
+    Eigen::internal::aligned_free(_data);
+    Eigen::internal::aligned_free(_quantizedData);
 #else // TARGET_PI
     free(_data);
     free(_quantizedData);
diff --git a/source/src/lib/math/matrix_gemm.cpp b/source/src/lib/math/matrix_gemm.cpp
index 080b5dd..1ecc5a2 100644
--- a/source/src/lib/math/matrix_gemm.cpp
+++ b/source/src/lib/math/matrix_gemm.cpp
@@ -391,6 +391,9 @@ void cblas_sgemm_fixed(
   posix_memalign((void**)(&aSubMatrix), 16, bytesPerSubMatrix);
   jpfloat_t* cSubMatrix;
   posix_memalign((void**)(&cSubMatrix), 16, bytesPerSubMatrix);
+#elif defined(USE_EIGEN_GEMM)
+  jpfloat_t* aSubMatrix = (jpfloat_t*)(Eigen::internal::aligned_malloc(bytesPerSubMatrix));
+  jpfloat_t* cSubMatrix = (jpfloat_t*)(Eigen::internal::aligned_malloc(bytesPerSubMatrix));
 #else // TARGET_PI
   jpfloat_t* aSubMatrix = (jpfloat_t*)(malloc(bytesPerSubMatrix));
   jpfloat_t* cSubMatrix = (jpfloat_t*)(malloc(bytesPerSubMatrix));
@@ -576,8 +579,13 @@ void cblas_sgemm_fixed(
 #endif // USE_EIGEN_GEMM
   }
 
+#if defined(USE_EIGEN_GEMM)
+  Eigen::internal::aligned_free(aSubMatrix);
+  Eigen::internal::aligned_free(cSubMatrix);
+#else // USE_EIGEN_GEMM
   free(aSubMatrix);
   free(cSubMatrix);
+#endif // USE_EIGEN_GEMM
 }
 #endif
 
diff --git a/source/src/lib/pi/qpu_gemm.cpp b/source/src/lib/pi/qpu_gemm.cpp
index bbbf0fb..da42cf0 100644
--- a/source/src/lib/pi/qpu_gemm.cpp
+++ b/source/src/lib/pi/qpu_gemm.cpp
@@ -1,3 +1,4 @@
+#ifdef USE_QPU_GEMM
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -13,7 +14,7 @@
 #define NUM_QPUS        (12)
 #define NUM_MESSAGE_VALUES (2)
 
-//#define DO_LOG_OPERATIONS
+
 
 extern uint32_t g_gemm_8bitCode[];
 extern size_t g_gemm_8bitCodeByteCount;
@@ -451,3 +452,5 @@ fprintf(stderr, "\n");
   delete outputGPU;
   delete input;
 }
+
+#endif // USE_QPU_GEMM