Add cuda convertfp16tofloat

src/Function1D.cu

@@ -1590,3 +1590,46 @@ CudaMatrix Aurora::uniqueByRows(const CudaMatrix& aMatrix, CudaMatrix& aIndexRes
     delete [] indexResult;
     return transpose(CudaMatrix::fromRawData(resultData, rows, columns));
 }
+
+__global__ void convertValueKernel(short* aSrc ,float* aDes, unsigned int size){
+    __shared__ ushort CONVERT_AND_VALUE;
+    __shared__ ushort CONVERT_AND_VALUE_2;
+    __shared__ ushort CONVERT_MUL_VALUE;
+    __shared__ unsigned int CONVERT_ADD_VALUE;
+    if (threadIdx.x == 0){
+        CONVERT_AND_VALUE = 15u;
+        CONVERT_AND_VALUE_2 = 2047u;
+        CONVERT_MUL_VALUE = 2048u;
+        CONVERT_ADD_VALUE = UINT32_MAX - 4095u;
+    }
+    __syncthreads();
+    unsigned int idx = blockIdx.x*blockDim.x +threadIdx.x;
+    if (idx >= size) return; 
+    short value = aSrc[idx];
+    ushort exponent=(ushort)value;
+    exponent = (exponent >> 11) & CONVERT_AND_VALUE;
+    short sign = value;
+    unsigned int sign_bit = (unsigned int)(sign < 0 ? 1 : 0);
+    ushort fraction3= (ushort)value;
+    fraction3 &= CONVERT_AND_VALUE_2;
+    unsigned int hidden_bit = sign_bit * (!exponent ? 1 : 0) * CONVERT_MUL_VALUE +
+                      ((!sign_bit && exponent) ? 1 : 0) * CONVERT_MUL_VALUE;
+    unsigned int temp = fraction3 + hidden_bit + sign_bit * CONVERT_ADD_VALUE;
+    int ret = (int)(exponent> 1 ? (temp << (exponent - 1)): temp);
+    aDes[idx] = (float)ret;
+}
+
+CudaMatrix Aurora::convertfp16tofloatCuda(short* aData, int aRows, int aColumns)
+{
+    unsigned int size = aRows*aColumns;
+    unsigned int short_size = size*sizeof(short);
+    short* input = nullptr;
+    cudaMalloc((void**)&input, short_size);
+    cudaMemcpy(input, aData, short_size, cudaMemcpyHostToDevice);
+    //uint16变换为float(32位)输出大小翻倍
+    float* output = nullptr;
+    cudaMalloc((void**)&output,size*sizeof(float));
+    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+    convertValueKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(input,output, size);
+    return CudaMatrix::fromRawData(output, aRows, aColumns);
+}