phi::MixVector<size_t> mixv_batch_indices(&batch_indices);

size_t *idx = mixv_batch_indices.CUDAMutableData(dev_ctx.GetPlace());

phi::funcs::SetConstant<phi::GPUContext, T> zero;

if (dFilter) {

T *dfilter = dev_ctx.template Alloc<T>(dFilter);

if (future_context <= 32) {

dim3 block_dim = dim3(32, 32);

future_context * block_y) *

sizeof(T);

RowConvGradFilterImproved<T>

in,

dout,

num_sequence,

int mem_per_block =

(block_x * block_y * 2) * sizeof(T); // For 2 arrays of size 32x32

RowConvGradFilter<T>

in,

dout,

num_sequence,

dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);

int mem_per_block = (future_context * block_dim.x) * sizeof(T);

RowConvGradInputSharedMemory<T>

dout, weights, num_sequence, input_dim, future_context, idx, din);

} else {

dim3 block_dim = dim3(32, 32);

dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);

dout, weights, num_sequence, input_dim, future_context, idx, din);

}

}