Merge pull request #341 from deephealthproject/develop

Add RiscV flags
deephealthproject · Oct 24, 2022 · d708158 · d708158
2 parents 5ad2a2a + 6964bca
commit d708158
Show file tree

Hide file tree

Showing 24 changed files with 117 additions and 55 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -7,6 +7,7 @@ cmake_minimum_required(VERSION 3.9.2)
 
 # User options
 option(BUILD_SUPERBUILD "Compile using the superbuild system" OFF)
+option(BUILD_RISCV "Compile for a RISCV cpu" OFF)
 option(BUILD_PROTOBUF "Compile using Protobuf" ON)
 option(BUILD_OPENMP "Compile using OpenMP" ON)
 option(BUILD_HPC "Compile using aggressive flags for performance" ON)
@@ -194,4 +195,4 @@ endif()
 if(BUILD_TESTS AND BUILD_HPC)
     message(WARNING "[WARNING] Some logic functions are not compatible with the 'BUILD_HPC' flag.
 If you're unit testing, or using one of these logic functions: isfinite(), isinf(), isnan(), isposinf(), isneginf(); then we recommend you to disable the HPC flag: '-D BUILD_HPC=OFF' to obtain the expected results.")
-endif()
+endif()
diff --git a/Jenkinsfile b/Jenkinsfile
@@ -7,15 +7,15 @@ pipeline {
                     agent {
                         docker { 
                             label 'docker'
-                            image 'pritt/base'
+                            image 'aimagelab/base'
                         }
                     }
                     stages {
                         stage('Build') {
                             steps {
                                 timeout(60) {
                                     echo 'Building..'
-                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=CPU -DBUILD_SUPERBUILD=ON -DBUILD_TESTS=ON', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
+                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=CPU -DBUILD_SUPERBUILD=ON -DBUILD_TESTS=ON -DBUILD_HPC=OFF -DBUILD_DIST=OFF -DBUILD_RUNTIME=OFF', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
                                         [args: '--parallel 4', withCmake: true]
                                     ]
                                 }
@@ -45,7 +45,7 @@ pipeline {
                             steps {
                                 timeout(60) {
                                     echo 'Building..'
-                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=CPU -DBUILD_SHARED_LIBS=OFF -DBUILD_SUPERBUILD=ON -DBUILD_TESTS=ON',  installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
+                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=CPU -DBUILD_SHARED_LIBS=OFF -DBUILD_SUPERBUILD=ON -DBUILD_TESTS=ON -DBUILD_HPC=OFF -DBUILD_DIST=OFF -DBUILD_RUNTIME=OFF',  installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
                                         [args: '--config Release --parallel 4', withCmake: true]
                                     ]
                                 }
@@ -70,7 +70,7 @@ pipeline {
                     agent {
                         docker { 
                             label 'docker && gpu'
-                            image 'pritt/base-cuda'
+                            image 'aimagelab/base-cuda'
                             args '--gpus 1'
                         }
                     }
@@ -79,7 +79,7 @@ pipeline {
                             steps {
                                 timeout(60) {
                                     echo 'Building..'
-                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=GPU -DBUILD_TESTS=ON -DBUILD_SUPERBUILD=ON', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
+                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=GPU -DBUILD_TESTS=ON -DBUILD_SUPERBUILD=ON -DBUILD_HPC=OFF -DBUILD_DIST=OFF -DBUILD_RUNTIME=OFF', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
                                         [args: '--parallel 4', withCmake: true]
                                     ]
                                 }
@@ -109,7 +109,7 @@ pipeline {
                             steps {
                                 timeout(60) {
                                     echo 'Building..'
-                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=GPU -DBUILD_TESTS=ON -DBUILD_SUPERBUILD=ON', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
+                                    cmakeBuild buildDir: 'build', buildType: 'Release', cmakeArgs: '-DBUILD_TARGET=GPU -DBUILD_TESTS=ON -DBUILD_SUPERBUILD=ON -DBUILD_HPC=OFF -DBUILD_DIST=OFF -DBUILD_RUNTIME=OFF', installation: 'InSearchPath', sourceDir: '.', cleanBuild: true, steps: [
                                         [args: '--config Release --parallel 4', withCmake: true]
                                     ]
                                 }

diff --git a/cmake/googletest.CMakeLists.txt.in b/cmake/googletest.CMakeLists.txt.in
@@ -10,7 +10,7 @@ include(ExternalProject)
 ExternalProject_Add(googletest
         PREFIX googletest
         GIT_REPOSITORY https://github.com/google/googletest.git
-        GIT_TAG "release-1.10.0"
+        GIT_TAG "release-1.11.0"
         SOURCE_DIR "${EP_BASE_DIR}/googletest-src"
         BINARY_DIR "${EP_BASE_DIR}/googletest-build"
         CMAKE_CACHE_ARGS

diff --git a/examples/nn/1_mnist/1_mnist_mlp.cpp b/examples/nn/1_mnist/1_mnist_mlp.cpp
@@ -36,7 +36,7 @@ int main(int argc, char **argv) {
 
     // Settings
     int epochs = (testing) ? 2 : 10;
-    int batch_size = 200;
+    int batch_size = 100;
     int num_classes = 10;
 
     // Define network

diff --git a/examples/nn/1_mnist/7_mnist_conv.cpp b/examples/nn/1_mnist/7_mnist_conv.cpp
@@ -23,20 +23,33 @@ using namespace eddl;
 //////////////////////////////////
 
 int main(int argc, char **argv) {
+
+    // Settings
+    int epochs = 10;
+    int batch_size = 100;
+    int num_classes = 10;
     bool testing = false;
     bool use_cpu = false;
+    int num_cpu_threads = -1; // -1 = the maximum available cores
+
     for (int i = 1; i < argc; ++i) {
-        if (strcmp(argv[i], "--testing") == 0) testing = true;
-        else if (strcmp(argv[i], "--cpu") == 0) use_cpu = true;
+        if (strcmp(argv[i], "--testing") == 0) {
+            testing = true;
+            epochs = 2;
+        } else if (strcmp(argv[i], "--cpu") == 0) {
+            use_cpu = true;
+        } else if (strcmp(argv[i], "--threads") == 0) {
+            num_cpu_threads = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--epochs") == 0) {
+            epochs = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--batch-size") == 0) {
+            batch_size = atoi(argv[++i]);
+        }
     }
 
     // Download mnist
     download_mnist();
 
-    // Settings
-    int epochs = (testing) ? 2 : 10;
-    int batch_size = 100;
-    int num_classes = 10;
 
     // Define network
     layer in = Input({784});
@@ -62,7 +75,7 @@ int main(int argc, char **argv) {
 
     compserv cs = nullptr;
     if (use_cpu) {
-        cs = CS_CPU();
+        cs = CS_CPU(num_cpu_threads);
     } else {
         cs = CS_GPU({1}, "full_mem"); // one GPU
         // cs = CS_GPU({1,1},100); // two GPU with weight sync every 100 batches

diff --git a/formulas/brew/eddl.rb b/formulas/brew/eddl.rb
@@ -4,8 +4,8 @@
 class Eddl < Formula
   desc "European Distributed Deep Learning Library (EDDL)"
   homepage "https://github.com/deephealthproject/eddl"
-  url "https://github.com/deephealthproject/eddl/archive/v1.0.2a.tar.gz"
-  sha256 "70d6067d44cb547e218236e5bb72faf45f602a7b8214f869626ca47f241481a0"
+  url "https://github.com/deephealthproject/eddl/archive/v1.1b.tar.gz"
+  sha256 "b9fe2bdc63808ae8e1a8eec96f66106c49b7a5ce9ee32ffe17fd6cf9d1b2c4ec"
 
   depends_on "cmake" => :build
   depends_on "eigen" => :build

diff --git a/formulas/conda/eddl/build.sh b/formulas/conda/eddl/build.sh
@@ -30,7 +30,7 @@ echo "CPU_COUNT=$CPU_COUNT"
 echo "#################################################"
 
 # Build makefiles
-mkdir build && cd build/ && cmake .. -DBUILD_TARGET=CPU \
+mkdir build && cd build/ && cmake .. -DBUILD_TARGET=CUDNN \
       -DBUILD_SUPERBUILD=OFF \
       -DBUILD_EXAMPLES=OFF \
       -DBUILD_TESTS=OFF \

diff --git a/formulas/conda/eddl/meta.yaml b/formulas/conda/eddl/meta.yaml
@@ -1,6 +1,6 @@
-{% set name = "eddl-cpu" %}  # eddl-cpu, eddl-gpu, eddl-cudnn
+{% set name = "eddl-cudnn" %}  # eddl-cpu, eddl-gpu, eddl-cudnn
 {% set version = "1.1b" %}
-{% set sha256 = "534ca4bd5e83236c164a26a058e08df971d28c0ae5b32674bec364ae0370843a" %}
+{% set sha256 = "b9fe2bdc63808ae8e1a8eec96f66106c49b7a5ce9ee32ffe17fd6cf9d1b2c4ec" %}
 
 package:
   name: {{ name|lower }}

diff --git a/scripts/install/install_gtest.sh b/scripts/install/install_gtest.sh
@@ -5,14 +5,17 @@ sudo apt-get update
 sudo apt-get wget
 
 # Build and install
-wget https://github.com/google/googletest/archive/release-1.10.0.tar.gz
-tar -zxvf release-1.10.0.tar.gz
-cd googletest-release-1.10.0  && mkdir build && \
+#wget https://github.com/google/googletest/archive/release-1.10.0.tar.gz
+#tar -zxvf release-1.10.0.tar.gz
+#cd googletest-release-1.10.0  && mkdir build && \
+wget https://github.com/google/googletest/archive/release-1.11.0.tar.gz
+tar -zxvf release-1.11.0.tar.gz
+cd googletest-release-1.11.0  && mkdir build && \
 cd build && cmake .. && sudo make -j$(nproc) install
 
 # Build and install
 #sudo apt-get install -y libgtest-dev
 #cd /usr/src/gtest
 #cmake CMakeLists.txt
 #make -j$(nproc)
-#cp *.a /usr/lib
+#cp *.a /usr/lib
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -12,6 +12,7 @@ set_property(CACHE BUILD_TARGET PROPERTY STRINGS CPU GPU CUDNN FPGA)
 
 # Initializations (Local's scope)
 SET(USE_OPENMP OFF)
+SET(USE_RISCV OFF)
 SET(USE_CUDA OFF)
 SET(USE_CUDNN OFF)
 SET(USE_FPGA OFF)
@@ -20,6 +21,8 @@ SET(USE_PROTOBUF OFF)
 # Device specific setup
 string(TOUPPER ${BUILD_TARGET} BUILD_TARGET)  # Detect cpu, Cpu, CPU,...
 if(${BUILD_TARGET} STREQUAL "CPU")
+elseif(${BUILD_TARGET} STREQUAL "RISCV")
+    SET(USE_RISCV ON) # Local's scope
 elseif(${BUILD_TARGET} STREQUAL "GPU" OR ${BUILD_TARGET} STREQUAL "CUDA")
     SET(USE_CUDA ON) # Local's scope
 elseif(${BUILD_TARGET} STREQUAL "CUDNN")
@@ -122,6 +125,11 @@ endif()
 ########################## COMPILE FOR RISK-V #############################
 ###########################################################################
 
+if(USE_RISCV)
+    set(CMAKE_C_FLAGS_RELEASE "-O2 -march=rv64imafdc -mabi=lp64d ")
+    set(CMAKE_CXX_FLAGS_RELEASE "-O2 -march=rv64imafdc -mabi=lp64d ")
+endif()
+
 if(RISKV_COMPILER_PATH)
     message(STATUS "Compiling for Risk-V...")
 endif()
@@ -499,10 +507,8 @@ if(USE_FPGA)
     message(STATUS "OpenCL include: " ${OpenCL_INCLUDE_DIRS} )
     message(STATUS "OpenCL libraries: " ${OpenCL_LIBRARIES} )
 endif()
-if(RISKV_COMPILER_PATH)
-    message(STATUS "-------------------------------------------" )
-    message(STATUS "Risk-V compiler: " ${RISKV_COMPILER_PATH} )
-endif()
+message(STATUS "-------------------------------------------" )
+message(STATUS "RISCV enabled: " ${USE_RISCV} )
 message(STATUS "-------------------------------------------" )
 message(STATUS "Eigen3 root dir: " ${Eigen3_DIR} )
 message(STATUS "Eigen3 include: " ${EIGEN3_INCLUDE_DIR} )

diff --git a/src/descriptors/descriptor_conv2D.cpp b/src/descriptors/descriptor_conv2D.cpp
@@ -77,6 +77,9 @@ ConvolDescriptor::~ConvolDescriptor(){
             delete gpuOB;
         }
     }
+#else
+    cudnnDestroyTensorDescriptor(xDesc);
+    cudnnDestroyTensorDescriptor(yDesc);
 #endif
 #endif
 
@@ -219,8 +222,7 @@ void ConvolDescriptor::build(Tensor *A) {
                                     convolution_mode, data_type);
 
    cudnnCreateTensorDescriptor(&xDesc);
-   cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type,
-                 in,iz,ir,ic);
+   cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type, in,iz,ir,ic);
 
    cudnnCreateFilterDescriptor(&wDesc);
    cudnnSetFilter4dDescriptor(wDesc, data_type, tensor_format, nk, kz, kr, kc);
@@ -264,11 +266,15 @@ void ConvolDescriptor::resize(int b)
 #endif
 
 #ifdef cCUDNN
-   cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type,
-                 b,iz,ir,ic);
+   cudnnDestroyTensorDescriptor(xDesc);
+   cudnnDestroyTensorDescriptor(yDesc);
 
+   cudnnCreateTensorDescriptor(&xDesc);
    cudnnCreateTensorDescriptor(&yDesc);
-   cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, O->shape[0], O->shape[1],O->shape[2],O->shape[3]);
+
+   cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type, b,iz,ir,ic);
+   cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, b, O->shape[1],O->shape[2],O->shape[3]);
+
    cudnn_env_init = -1;
    cudnn_conv_back_init = -1;
 

diff --git a/src/descriptors/descriptor_conv3D.cpp b/src/descriptors/descriptor_conv3D.cpp
@@ -293,15 +293,19 @@ void ConvolDescriptor3D::resize(int b)
 #endif
 
 #ifdef cCUDNN
+    cudnnDestroyTensorDescriptor(xDesc);
+    cudnnDestroyTensorDescriptor(yDesc);
+
+    cudnnCreateTensorDescriptor(&xDesc);
+    cudnnCreateTensorDescriptor(&yDesc);
+
    int dims[5] = {b, iz, id, ir, ic};
    int str[5] = {iz*id*ir*ic,id*ir*ic,ir*ic,ic,1};
-   cudnnSetTensorNdDescriptor(xDesc, /*tensor_format,*/ data_type,5,dims,str);
+    cudnnSetTensorNdDescriptor(xDesc, /*tensor_format,*/ data_type,5,dims,str);
 
    int ydims[5] = {b,z,d,r,c};
    int ystr[5] = {z*d*r*c, d*r*c, r*c, c, 1};
-   cudnnSetTensorNdDescriptor(yDesc, /*tensor_format,*/ data_type, 5, ydims, ystr);
-
-   //cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, O->shape[0], O->shape[1],O->shape[2],O->shape[3]);
+    cudnnSetTensorNdDescriptor(yDesc, /*tensor_format,*/ data_type, 5, ydims, ystr);
 
    cudnn_env_init = -1;
    cudnn_conv_back_init = -1;

diff --git a/src/descriptors/descriptor_pool2D.cpp b/src/descriptors/descriptor_pool2D.cpp
@@ -131,8 +131,7 @@ void PoolDescriptor::build(Tensor *A) {
     tensor_format = CUDNN_TENSOR_NCHW;  // CUDNN_TENSOR_NHWC
 
     cudnnCreateTensorDescriptor(&xDesc);
-    cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type,
-                 in,iz,ir,ic);
+    cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type, in,iz,ir,ic);
     cudnnCreateTensorDescriptor(&yDesc);
     cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, in, z,r,c);
 }
@@ -146,10 +145,14 @@ void PoolDescriptor::resize(int b) {
   this->O->resize(b);
 #ifdef cCUDNN
     if(!this->O->isCPU()){
-   cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type,
-                 b,iz,ir,ic);
-   cudnnCreateTensorDescriptor(&yDesc);
-   cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, O->shape[0], O->shape[1],O->shape[2],O->shape[3]);
+        cudnnDestroyTensorDescriptor(xDesc);
+        cudnnDestroyTensorDescriptor(yDesc);
+
+        cudnnCreateTensorDescriptor(&xDesc);
+        cudnnCreateTensorDescriptor(&yDesc);
+
+       cudnnSetTensor4dDescriptor(xDesc, tensor_format, data_type, b,iz,ir,ic);
+       cudnnSetTensor4dDescriptor(yDesc, tensor_format, data_type, b, O->shape[1],O->shape[2],O->shape[3]);
 }
 #endif
 //  if (!mem_level) { D->resize(b); }

diff --git a/src/descriptors/descriptor_pool3D.cpp b/src/descriptors/descriptor_pool3D.cpp
@@ -167,14 +167,19 @@ void PoolDescriptor3D::resize(int b) {
   O->resize(b);
 #ifdef cCUDNN
   if(!O->isCPU()){
-   int dims[5] = {b, iz, id, ir, ic};
-   int str[5] = {iz*id*ir*ic,id*ir*ic,ir*ic,ic,1};
-   cudnnSetTensorNdDescriptor(xDesc, data_type,5,dims,str);
+      cudnnDestroyTensorDescriptor(xDesc);
+      cudnnDestroyTensorDescriptor(yDesc);
 
-   int ydims[5] = {b,z,d,r,c};
-   int ystr[5] = {z*d*r*c, d*r*c, r*c, c, 1};
-   cudnnSetTensorNdDescriptor(yDesc, data_type, 5, ydims, ystr);
+      cudnnCreateTensorDescriptor(&xDesc);
+      cudnnCreateTensorDescriptor(&yDesc);
+
+       int dims[5] = {b, iz, id, ir, ic};
+       int str[5] = {iz*id*ir*ic,id*ir*ic,ir*ic,ic,1};
+       cudnnSetTensorNdDescriptor(xDesc, data_type,5,dims,str);
 
+       int ydims[5] = {b,z,d,r,c};
+       int ystr[5] = {z*d*r*c, d*r*c, r*c, c, 1};
+       cudnnSetTensorNdDescriptor(yDesc, data_type, 5, ydims, ystr);
 }
 
 

diff --git a/src/layers/conv/layer_conv1D.cpp b/src/layers/conv/layer_conv1D.cpp
@@ -112,6 +112,8 @@ void LConv1D::mem_delta(){
         if(this->verbosity_level >= 2) {
             std::cout << "Booked delta for: " + this->name << std::endl;
         }
+    } else {
+        this->delta->resize(this->output->shape[0]);
     }
 }
 

diff --git a/src/layers/conv/layer_conv2D.cpp b/src/layers/conv/layer_conv2D.cpp
@@ -83,6 +83,8 @@ void LConv::mem_delta(){
         if(this->verbosity_level >= 2) {
             std::cout << "Booked delta for: " + this->name << std::endl;
         }
+    } else {
+        this->delta->resize(this->output->shape[0]);
     }
 }
 

diff --git a/src/layers/conv/layer_conv3D.cpp b/src/layers/conv/layer_conv3D.cpp
@@ -89,6 +89,8 @@ void LConv3D::mem_delta(){
         if(this->verbosity_level >= 2) {
             std::cout << "Booked delta for: " + this->name << std::endl;
         }
+    } else {
+        this->delta->resize(this->output->shape[0]);
     }
 }
 

diff --git a/src/layers/pool/layer_pool1D.cpp b/src/layers/pool/layer_pool1D.cpp
@@ -61,6 +61,8 @@ void LPool1D::mem_delta(){
         if(this->verbosity_level >= 2) {
             std::cout << "Booked delta for: " + this->name << std::endl;
         }
+    } else {
+        this->delta->resize(this->output->shape[0]);
     }
 }
 

diff --git a/src/layers/pool/layer_pool2D.cpp b/src/layers/pool/layer_pool2D.cpp
@@ -52,6 +52,8 @@ void LPool::mem_delta(){
         if(this->verbosity_level >= 2) {
             std::cout << "Booked delta for: " + this->name << std::endl;
         }
+    } else {
+        this->delta->resize(this->output->shape[0]);
     }
 }