[build,src] Enhancements to the cudamatrix/cudavector classes. (#3373)

* Added CuSolver to the matrix class.  This is only supported with
Cuda 9.1 or newer.  Calling CuSolver code without Cuda 9.1 or newer
will result in a runtime error.

This change required some changes to the build system which requires
versioning the configure script. This forces everyone to reconfigure.
Failure to reconfigure would result in linking and build errors on
some systems.

src/configure

@@ -39,7 +39,7 @@
 
 # This should be incremented after any significant change to the configure
 # script, i.e. any change affecting kaldi.mk or the build system as a whole.
-CONFIGURE_VERSION=10
+CONFIGURE_VERSION=11
 
 # We support bash version 3.2 (Macs still ship with this version as of 2019)
 # and above.
@@ -433,22 +433,32 @@ function configure_cuda {
         7_*)
           MIN_UNSUPPORTED_GCC_VER="5.0"
           MIN_UNSUPPORTED_GCC_VER_NUM=50000;
+          CUSOLVER=false
         ;;
         8_*)
           MIN_UNSUPPORTED_GCC_VER="6.0"
           MIN_UNSUPPORTED_GCC_VER_NUM=60000;
+          CUSOLVER=false
         ;;
-        9_0 | 9_1)
+        9_0)
           MIN_UNSUPPORTED_GCC_VER="7.0"
           MIN_UNSUPPORTED_GCC_VER_NUM=70000;
+          CUSOLVER=false
+        ;;
+        9_1)
+          MIN_UNSUPPORTED_GCC_VER="7.0"
+          MIN_UNSUPPORTED_GCC_VER_NUM=70000;
+          CUSOLVER=true
         ;;
         9_2 | 9_* | 10_0)
           MIN_UNSUPPORTED_GCC_VER="8.0"
           MIN_UNSUPPORTED_GCC_VER_NUM=80000;
+          CUSOLVER=true
         ;;
         10_1 | 10_*)
           MIN_UNSUPPORTED_GCC_VER="9.0"
           MIN_UNSUPPORTED_GCC_VER_NUM=90000;
+          CUSOLVER=true
         ;;
         *)
           echo "Unsupported CUDA_VERSION (CUDA_VERSION=$CUDA_VERSION), please report it to Kaldi mailing list, together with 'nvcc -h' or 'ptxas -h' which lists allowed -gencode values..."; exit 1;
@@ -492,6 +502,8 @@ function configure_cuda {
     echo CUDA = true >> kaldi.mk
     echo CUDATKDIR = $CUDATKDIR >> kaldi.mk
     echo "CUDA_ARCH = $CUDA_ARCH" >> kaldi.mk
+
+
     echo >> kaldi.mk
 
     # 64bit/32bit? We do not support cross compilation with CUDA so, use direct
@@ -512,6 +524,11 @@ WARNING: CUDA will not be used!
          CUDA is not supported with 32-bit builds."
       exit 1;
     fi
+
+    #add cusolver flags for newer toolkits
+    if [[ $CUSOLVER -eq true ]]; then
+      echo "CUDA_LDLIBS += -lcusolver" >> kaldi.mk
+    fi
 
   else
     echo "\

src/cudamatrix/cu-common.h

@@ -59,6 +59,15 @@
   } \
 }
 
+#define CUSOLVER_SAFE_CALL(fun) \
+{ \
+  int32 ret; \
+  if ((ret = (fun)) != 0) { \
+    KALDI_ERR << "cusolverStatus_t " << ret << " : \"" << ret << "\" returned from '" << #fun << "'"; \
+  } \
+}
+
+
 #define CUSPARSE_SAFE_CALL(fun) \
 { \
   int32 ret; \

src/cudamatrix/cu-device.cc

@@ -110,15 +110,21 @@ void CuDevice::Initialize() {
     // Initialize CUBLAS.
     CUBLAS_SAFE_CALL(cublasCreate(&cublas_handle_));
     CUBLAS_SAFE_CALL(cublasSetStream(cublas_handle_, cudaStreamPerThread));
+
+#if CUDA_VERSION >= 9100
+    CUSOLVER_SAFE_CALL(cusolverDnCreate(&cusolverdn_handle_));
+    CUSOLVER_SAFE_CALL(cusolverDnSetStream(cusolverdn_handle_, 
+            cudaStreamPerThread));
+#endif
 
-    #if CUDA_VERSION >= 9000 
+#if CUDA_VERSION >= 9000 
     if (device_options_.use_tensor_cores) {
       // Enable tensor cores in CUBLAS
       // Note if the device does not support tensor cores this will fall back to normal math mode
       CUBLAS_SAFE_CALL(cublasSetMathMode(cublas_handle_, 
             CUBLAS_TENSOR_OP_MATH));
     }
-    #endif
+#endif
 
     // Initialize the cuSPARSE library
     CUSPARSE_SAFE_CALL(cusparseCreate(&cusparse_handle_));
@@ -130,6 +136,7 @@ void CuDevice::Initialize() {
     // To get same random sequence, call srand() before the constructor is invoked,
     CURAND_SAFE_CALL(curandSetGeneratorOrdering(
           curand_handle_, CURAND_ORDERING_PSEUDO_DEFAULT));
+    CURAND_SAFE_CALL(curandSetStream(curand_handle_, cudaStreamPerThread));
     SeedGpu();
   }
 }
@@ -263,6 +270,23 @@ void CuDevice::FinalizeActiveGpu() {
     // Initialize CUBLAS.
     CUBLAS_SAFE_CALL(cublasCreate(&cublas_handle_));
     CUBLAS_SAFE_CALL(cublasSetStream(cublas_handle_, cudaStreamPerThread));
+
+#if CUDA_VERSION >= 9100 
+    CUSOLVER_SAFE_CALL(cusolverDnCreate(&cusolverdn_handle_));
+    CUSOLVER_SAFE_CALL(cusolverDnSetStream(cusolverdn_handle_,
+            cudaStreamPerThread));
+#endif
+
+#if CUDA_VERSION >= 9000 
+    if (device_options_.use_tensor_cores) {
+      // Enable tensor cores in CUBLAS
+      // Note if the device does not support tensor cores this will fall back to normal math mode
+      CUBLAS_SAFE_CALL(cublasSetMathMode(cublas_handle_, 
+            CUBLAS_TENSOR_OP_MATH));
+    }
+#endif
+
+
     // Initialize the cuSPARSE library
     CUSPARSE_SAFE_CALL(cusparseCreate(&cusparse_handle_));
     CUSPARSE_SAFE_CALL(cusparseSetStream(cusparse_handle_, cudaStreamPerThread));
@@ -537,7 +561,8 @@ CuDevice::CuDevice():
     initialized_(false),
     device_id_copy_(-1),
     cublas_handle_(NULL),
-    cusparse_handle_(NULL) {
+    cusparse_handle_(NULL),
+    cusolverdn_handle_(NULL) {
 }
 
 CuDevice::~CuDevice() {
@@ -548,6 +573,11 @@ CuDevice::~CuDevice() {
   if (curand_handle_) {
     CURAND_SAFE_CALL(curandDestroyGenerator(curand_handle_));
   }
+#if CUDA_VERSION >= 9100
+  if (cusolverdn_handle_) {
+    CUSOLVER_SAFE_CALL(cusolverDnDestroy(cusolverdn_handle_));
+  }
+#endif
 }
 
 

src/cudamatrix/cu-device.h

@@ -37,6 +37,16 @@
 #include "cudamatrix/cu-allocator.h"
 #include "cudamatrix/cu-common.h"
 
+#if CUDA_VERSION >= 9100
+#include <cusolverDn.h>
+#else
+// cusolver not supported.  
+// Setting a few types to minimize compiler guards.
+// If a user tries to use cusovler it will throw an error.
+typedef void* cusolverDnHandle_t;
+typedef int cusolverStatus_t;
+#endif
+
 namespace kaldi {
 
 class CuTimer;
@@ -83,6 +93,13 @@ class CuDevice {
   inline cublasHandle_t GetCublasHandle() { return cublas_handle_; }
   inline cusparseHandle_t GetCusparseHandle() { return cusparse_handle_; }
   inline curandGenerator_t GetCurandHandle() { return curand_handle_; }
+  inline cusolverDnHandle_t GetCusolverDnHandle() { 
+#if CUDA_VERSION < 9100
+    KALDI_ERR << "CUDA VERSION '" << CUDA_VERSION << "' not new enough to support "
+      << "cusolver. Upgrade to at least 9.1";
+#endif
+    return cusolverdn_handle_; 
+  }
 
   inline void SeedGpu() {
     if (CuDevice::Instantiate().Enabled()) {
@@ -304,6 +321,7 @@ class CuDevice {
   cublasHandle_t cublas_handle_;
   cusparseHandle_t cusparse_handle_;
   curandGenerator_t curand_handle_;
+  cusolverDnHandle_t cusolverdn_handle_;
 }; // class CuDevice
 
 
@@ -322,6 +340,10 @@ inline cublasHandle_t GetCublasHandle() {
   return CuDevice::Instantiate().GetCublasHandle(); 
 }
 
+inline cusolverDnHandle_t GetCusolverDnHandle() { 
+  return CuDevice::Instantiate().GetCusolverDnHandle(); 
+}
+
 // A more convenient way to get the handle to use cuSPARSE APIs.
 inline cusparseHandle_t GetCusparseHandle() { 
   return CuDevice::Instantiate().GetCusparseHandle();