Merge branch 'fused_dense_develop' into 'master'

fix revert fused_dense to fp32_r

See merge request aicomponent/apex!4

csrc/fused_dense_cuda.cu

@@ -163,53 +163,6 @@ cublasStatus_t gemm_bias(
     const float* beta,
     at::Half* C,
     int ldc) {
-<<<<<<< HEAD
-  if (parseEnvVarFlag("APEX_ROCBLAS_GEMM_ALLOW_HALF")) {
-    half h_alpha = __float2half(*alpha);
-    half h_beta = __float2half(*beta);
-    return cublasGemmEx(
-        handle,
-        transa,
-        transb,
-        m,
-        n,
-        k,
-        /* alpha */ &h_alpha,
-        A,
-        CUDA_R_16F,
-        lda,
-        B,
-        CUDA_R_16F,
-        ldb,
-        /* beta */ &h_beta,
-        C,
-        CUDA_R_16F,
-        ldc,
-        /* CUDA_R_32F */ CUDA_R_16F,
-        CUBLAS_GEMM_DEFAULT_TENSOR_OP);
-  } else {
-    return cublasGemmEx(
-        handle,
-        transa,
-        transb,
-        m,
-        n,
-        k,
-        alpha,
-        A,
-        CUDA_R_16F,
-        lda,
-        B,
-        CUDA_R_16F,
-        ldb,
-        beta,
-        C,
-        CUDA_R_16F,
-        ldc,
-        CUDA_R_32F,
-        CUBLAS_GEMM_DEFAULT_TENSOR_OP);
-  }
-=======
 #ifdef __HIP_PLATFORM_HCC__
   return rocblas_gemm_ex(
       handle,
@@ -258,7 +211,6 @@ cublasStatus_t gemm_bias(
       CUDA_R_32F,
       CUBLAS_GEMM_DEFAULT_TENSOR_OP);
 #endif
->>>>>>> mirror/master
 }