/// indicate that scalar values are passed as reference on device
inline void set_scalar_host()
{
MGPU_CUDA_BLAS_CALL(cublasSetPointerMode(handle_,
CUBLAS_POINTER_MODE_HOST));
}
static int setup(void *c) {
cuda_context *ctx = (cuda_context *)c;
blas_handle *handle;
const char *tmp[2];
cublasStatus_t err;
int e;
int types[10];
if (ctx->blas_handle != NULL)
return GA_NO_ERROR;
handle = calloc(1, sizeof(*handle));
if (handle == NULL)
return GA_MEMORY_ERROR;
cuda_enter(ctx);
err = cublasCreate(&handle->h);
if (err != CUBLAS_STATUS_SUCCESS) {
cuda_exit(ctx);
free(handle);
return GA_BLAS_ERROR;
}
err = cublasSetStream(handle->h, ctx->s);
if (err != CUBLAS_STATUS_SUCCESS) {
e = GA_BLAS_ERROR;
goto e1;
}
cublasSetPointerMode(handle->h, CUBLAS_POINTER_MODE_HOST);
cublasSetAtomicsMode(handle->h, CUBLAS_ATOMICS_ALLOWED);
types[0] = GA_BUFFER;
types[1] = GA_SIZE;
types[2] = GA_BUFFER;
types[3] = GA_SIZE;
types[4] = GA_BUFFER;
types[5] = GA_SIZE;
types[6] = GA_SIZE;
types[7] = GA_SIZE;
types[8] = GA_SIZE;
e = GpuKernel_init(&handle->sgemvBH_N_a1_b1_small, &cuda_ops, ctx, 1, &code_sgemvBH_N_a1_b1_small, NULL, "sgemv", 9, types, 0, NULL);
if (e != GA_NO_ERROR) goto e1;
e = GpuKernel_init(&handle->sgemvBH_T_a1_b1_small, &cuda_ops, ctx, 1, &code_sgemvBH_T_a1_b1_small, NULL, "sgemv", 9, types, 0, NULL);
if (e != GA_NO_ERROR) goto e2;
tmp[0] = atomicadd_double;
tmp[1] = code_dgemvBH_N_a1_b1_small;
e = GpuKernel_init(&handle->dgemvBH_N_a1_b1_small, &cuda_ops, ctx, 2, tmp, NULL, "dgemv", 9, types, GA_USE_DOUBLE, NULL);
if (e != GA_NO_ERROR) goto e3;
tmp[0] = atomicadd_double;
tmp[1] = code_dgemvBH_T_a1_b1_small;
e = GpuKernel_init(&handle->dgemvBH_T_a1_b1_small, &cuda_ops, ctx, 2, tmp, NULL, "dgemv", 9, types, GA_USE_DOUBLE, NULL);
if (e != GA_NO_ERROR) goto e4;
types[0] = GA_BUFFER;
types[1] = GA_SIZE;
types[2] = GA_BUFFER;
types[3] = GA_SIZE;
types[4] = GA_FLOAT;
types[5] = GA_BUFFER;
types[6] = GA_SIZE;
types[7] = GA_SIZE;
types[8] = GA_SIZE;
types[9] = GA_SIZE;
e = GpuKernel_init(&handle->sgerBH_gen_small, &cuda_ops, ctx, 1, &code_sgerBH_gen_small, NULL, "_sgerBH_gen_small", 10, types, 0, NULL);
if (e != GA_NO_ERROR) goto e5;
types[4] = GA_DOUBLE;
tmp[0] = atomicadd_double;
tmp[1] = code_dgerBH_gen_small;
e = GpuKernel_init(&handle->dgerBH_gen_small, &cuda_ops, ctx, 2, tmp, NULL, "_dgerBH_gen_small", 10, types, GA_USE_DOUBLE, NULL);
if (e != GA_NO_ERROR) goto e6;
ctx->blas_handle = handle;
cuda_exit(ctx);
return GA_NO_ERROR;
e6:
GpuKernel_clear(&handle->sgerBH_gen_small);
e5:
GpuKernel_clear(&handle->dgemvBH_T_a1_b1_small);
e4:
GpuKernel_clear(&handle->dgemvBH_N_a1_b1_small);
e3:
GpuKernel_clear(&handle->sgemvBH_T_a1_b1_small);
e2:
GpuKernel_clear(&handle->sgemvBH_N_a1_b1_small);
e1:
cublasDestroy(handle->h);
cuda_exit(ctx);
free(handle);
return e;
}
/// indicate that scalar values are passed as reference on host
inline void set_scalar_device()
{
MGPU_CUDA_BLAS_CALL(cublasSetPointerMode(handle_,
CUBLAS_POINTER_MODE_DEVICE));
}
