Reproduce Cudnn RNN error

readme.md

make && ./RNN

got error: cuDNN Error: CUDNN_STATUS_BAD_PARAM RNN_example.cu 136

Makefile

# Location of the CUDA Toolkit
CUDA_PATH ?= /usr/local/cuda

# architecture
HOST_ARCH   := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH)

# Adjust this for ARMv7 with a 32-bit filesystem
ifeq ($(TARGET_ARCH), aarch64)
    ifeq ($(shell file /sbin/init | grep 32-bit), 1)
        TARGET_ARCH=armv7l
    endif
endif
 
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
    ifneq ($(TARGET_ARCH),$(HOST_ARCH))
        ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
            TARGET_SIZE := 64
        else ifneq (,$(filter $(TARGET_ARCH),armv7l))
            TARGET_SIZE := 32
        endif
    else
        TARGET_SIZE := $(shell getconf LONG_BIT)
    endif
else
    $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
    ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le))
        $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
    endif
endif

# operating system
HOST_OS   := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
TARGET_OS ?= $(HOST_OS)

ifeq ($(TARGET_OS),QNX)
TARGET_OS := qnx
endif

ifeq (,$(filter $(TARGET_OS),linux darwin qnx QNX android))
    $(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
endif

# host compiler
ifeq ($(TARGET_OS),darwin)
    ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
        HOST_COMPILER ?= clang++
    endif
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
    ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
        ifeq ($(TARGET_OS),linux)
            HOST_COMPILER ?= arm-linux-gnueabihf-g++
        else ifeq ($(TARGET_OS),qnx)
            ifeq ($(QNX_HOST),)
                $(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
            endif
            ifeq ($(QNX_TARGET),)
                $(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
            endif
            export QNX_HOST
            export QNX_TARGET
            HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
        else ifeq ($(TARGET_OS),android)
            HOST_COMPILER ?= arm-linux-androideabi-g++
        endif
    else ifeq ($(TARGET_ARCH),aarch64)
        ifeq ($(TARGET_OS), linux)
            HOST_COMPILER ?= aarch64-linux-gnu-g++
        else ifeq ($(TARGET_OS), android)
            HOST_COMPILER ?= aarch64-linux-android-g++
        endif
    else ifeq ($(TARGET_ARCH),ppc64le)
        HOST_COMPILER ?= powerpc64le-linux-gnu-g++
    endif
endif
HOST_COMPILER ?= g++
NVCC          := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)

# internal flags
NVCCFLAGS   := -m${TARGET_SIZE}
CCFLAGS     :=
LDFLAGS     :=

# build flags
ifeq ($(TARGET_OS),darwin)
    LDFLAGS += -rpath $(CUDA_PATH)/lib
    CCFLAGS += -arch $(HOST_ARCH)
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
    LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
    CCFLAGS += -mfloat-abi=hard
else ifeq ($(TARGET_OS),android)
    LDFLAGS += -pie
    CCFLAGS += -fpie -fpic -fexceptions
endif

ifneq ($(TARGET_ARCH),$(HOST_ARCH))
    ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
        ifneq ($(TARGET_FS),)
            GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
            ifeq ($(GCCVERSIONLTEQ46),1)
                CCFLAGS += --sysroot=$(TARGET_FS)
            endif
            LDFLAGS += --sysroot=$(TARGET_FS)
            LDFLAGS += -rpath-link=$(TARGET_FS)/lib
            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
        endif
    endif
endif

# Debug build flags
ifeq ($(dbg),1)
      NVCCFLAGS += -g -G
      BUILD_TYPE := debug
else
      BUILD_TYPE := release
endif

ALL_CCFLAGS :=
ALL_CCFLAGS += $(NVCCFLAGS)
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))

SAMPLE_ENABLED := 1

ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))

# Common includes and paths for CUDA
ifneq ($(TARGET_ARCH), ppc64le)
INCLUDES := -I$(CUDA_PATH)/include
else
INCLUDES := -I$(CUDA_PATH)/targets/ppc64le-linux/include
endif
LIBRARIES :=

################################################################################

# Gencode arguments
SMS ?= 30 35 50 53

ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
SAMPLE_ENABLED := 0
endif

ifeq ($(GENCODE_FLAGS),)
# Generate SASS code for each SM architecture listed in $(SMS)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))

# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
HIGHEST_SM := $(lastword $(sort $(SMS)))
ifneq ($(HIGHEST_SM),)
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif

INCLUDES += -I.
LIBRARIES += -L. -lcublas -lcudnn -lcudart -lstdc++ -lm

ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif

################################################################################

# Target rules
all: build

build: RNN

check.deps:
ifeq ($(SAMPLE_ENABLED),0)
	@echo "Sample will be waived due to the above missing dependencies"
else
	@echo "Sample is ready - all dependencies have been met"
endif

OBJ = RNN_example.o

RNN: $(OBJ)
	$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)

%.o: %.cu
	$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<

run: build
	$(EXEC) ./RNN 100 4 512 64 2

clean:
	rm -rf *o
	rm -rf RNN

clobber: clean

RNN_example.cu

#include <cudnn.h>
#include <cuda_fp16.h>
#include <cuda.h>
#include <stdio.h>

// Define some error checking macros.
#define cudaErrCheck(stat) { cudaErrCheck_((stat), __FILE__, __LINE__); }
void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
   if (stat != cudaSuccess) {
      fprintf(stderr, "CUDA Error: %s %s %d\n", cudaGetErrorString(stat), file, line);
   }
}

#define cudnnErrCheck(stat) { cudnnErrCheck_((stat), __FILE__, __LINE__); }
void cudnnErrCheck_(cudnnStatus_t stat, const char *file, int line) {
   if (stat != CUDNN_STATUS_SUCCESS) {
      fprintf(stderr, "cuDNN Error: %s %s %d\n", cudnnGetErrorString(stat), file, line);
   }
}

int main(int argc, char* argv[]) {

   // seqLength = 1
   int numLayers = 1;
   int hiddenSize = 128;
   int inputSize = 128;
   int miniBatch = 4;
   float dropout = 0;
   int persistent = 0;

   cudnnHandle_t cudnnHandle;
   cudnnErrCheck(cudnnCreate(&cudnnHandle));

   /// set up xDesc
   cudnnTensorDescriptor_t xDesc;
   cudnnErrCheck(cudnnCreateTensorDescriptor(&xDesc));
   int dimA[3];
   int strideA[3];
   dimA[0] = miniBatch;
   dimA[1] = inputSize;
   dimA[2] = 1;
   strideA[0] = dimA[2] * dimA[1];
   strideA[1] = dimA[2];
   strideA[2] = 1;
   cudnnErrCheck(cudnnSetTensorNdDescriptor(xDesc, CUDNN_DATA_HALF, 3, dimA, strideA));

   // -------------------------
   // Set up the dropout descriptor (needed for the RNN descriptor)
   // -------------------------
   unsigned long long seed = 1337ull; // Pick a seed.

   cudnnDropoutDescriptor_t dropoutDesc;
   cudnnErrCheck(cudnnCreateDropoutDescriptor(&dropoutDesc));

   // How much memory does dropout need for states?
   // These states are used to generate random numbers internally
   // and should not be freed until the RNN descriptor is no longer used
   size_t stateSize;
   void *states;
   cudnnErrCheck(cudnnDropoutGetStatesSize(cudnnHandle, &stateSize));
   cudaErrCheck(cudaMalloc(&states, stateSize));
   cudnnErrCheck(cudnnSetDropoutDescriptor(dropoutDesc,
                             cudnnHandle,
                             dropout,
                             states,
                             stateSize,
                             seed));

   // -------------------------
   // Set up the RNN descriptor
   // -------------------------
   cudnnRNNDescriptor_t rnnDesc;
   cudnnRNNMode_t RNNMode = CUDNN_LSTM;
   cudnnRNNAlgo_t RNNAlgo;
   cudnnErrCheck(cudnnCreateRNNDescriptor(&rnnDesc));

   // Persistent RNNs are only supported on Pascal+ GPUs.
   if      (persistent == 0) RNNAlgo = CUDNN_RNN_ALGO_STANDARD;
   else if (persistent == 1) RNNAlgo = CUDNN_RNN_ALGO_PERSIST_STATIC;
   else if (persistent == 2) RNNAlgo = CUDNN_RNN_ALGO_PERSIST_DYNAMIC;

   cudnnErrCheck(cudnnSetRNNDescriptor_v6(cudnnHandle,
                                       rnnDesc,
                                       hiddenSize,
                                       numLayers,
                                       dropoutDesc,
                                       CUDNN_LINEAR_INPUT, // We can also skip the input matrix transformation
                                       CUDNN_UNIDIRECTIONAL,
                                       RNNMode,
                                       RNNAlgo, // Can be changed to use persistent RNNs on Pascal+ GPUs.
                                       CUDNN_DATA_FLOAT)); // math precision

   cudnnErrCheck(cudnnSetRNNMatrixMathType(rnnDesc, CUDNN_TENSOR_OP_MATH));


   // -------------------------
   // Set up parameters
   // -------------------------
   // This needs to be done after the rnn descriptor is set as otherwise
   // we don't know how many parameters we have to allocate
   void *w;

   cudnnFilterDescriptor_t wDesc;
   cudnnErrCheck(cudnnCreateFilterDescriptor(&wDesc));

   size_t weightsSize;
   cudnnErrCheck(cudnnGetRNNParamsSize(cudnnHandle, rnnDesc, xDesc, &weightsSize, CUDNN_DATA_HALF));

   int dimW[3];
   dimW[0] =  weightsSize / sizeof(half);
   dimW[1] = 1;
   dimW[2] = 1;

   cudnnErrCheck(cudnnSetFilterNdDescriptor(wDesc, CUDNN_DATA_HALF, CUDNN_TENSOR_NCHW, 3, dimW));
   cudaErrCheck(cudaMalloc((void**)&w,  weightsSize));


   // -------------------------
   // Set up work space and reserved memory
   // -------------------------
   // Weights
   cudnnFilterDescriptor_t linLayerMatDesc;
   cudnnErrCheck(cudnnCreateFilterDescriptor(&linLayerMatDesc));
   half *linLayerMat;

   cudnnErrCheck(
           cudnnGetRNNLinLayerMatrixParams(
               cudnnHandle,
               rnnDesc,
               0,  /* layer*/
               xDesc,
               wDesc,
               w,
               0, /*linLayerID,*/
               linLayerMatDesc,
               (void**)&linLayerMat));
}