changed depthwise convolution

llvm/projects/hpvm-tensor-rt/tensor_runtime/include/approx_techniques.h

@@ -114,6 +114,7 @@ __global__ void depthwise_conv(float* const __restrict__ y,
 
         #pragma unroll
 	for (int i = 0; i < num; i++) {
+	  if(b + i < B)
 		y4d(b + i, m, 0, tx) = C[i];
 
 	}
@@ -176,7 +177,7 @@ void* tensorConvCutlass(void* input_ptr, void* filter_ptr,
 
 		}*/
 		
-		dim3 grid((n / 8), c);
+		dim3 grid((n / 8) + 1, c);
 		dim3 block(h * w);
 		depthwise_conv8 << <grid, block >> > ((float*)output->gpu_data,
 				(float*)input->gpu_data, (float*)filter->gpu_data,