CUDA tiled 2D Convolution in shared memory is slower than global memory - image-processing

Question

I performed two convolution using constant memory for mask.
One without tiling in global memory:
__global__ void constGradientConvolution(uint8_t* inputImgData, uint8_t* gradientImgData, int w, int h) {
// Calculate the global thread positions
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
// Starting index for calculation
int start_r = row - SOBEL_OP_RADIUS;
int start_c = col - SOBEL_OP_RADIUS;
// Temp value for calculation
int temp = 0;
// Iterate over all the rows
for (int i = 0; i < SOBEL_OP_DIM; i++) {
// Go over each column
for (int j = 0; j < SOBEL_OP_DIM; j++) {
// Range check for rows
if ((start_r + i) >= 0 && (start_r + i) < h) {
// Range check for columns
if ((start_c + j) >= 0 && (start_c + j) < w) {
// Accumulate result
temp += inputImgData[(start_r + i) * w + (start_c + j)] *
constMask[i * SOBEL_OP_DIM + j];
}
}
}
}
// Write back the result
gradientImgData[row * w + col] = (uint8_t)abs(temp);
}
and one with tiling, loading in shared memory, credits to https://www.cstechera.com/2015/07/two-dimensional-2d-image-convolution-in-CUDA.html:
__global__ void tiledGradientConvolution(uint8_t* inputImgData, uint8_t* gradientImgData, int width, int height) {
__shared__ uint8_t N_ds[SharedDim_y][SharedDim_x];
// First batch loading
int dest = threadIdx.y * TILE_WIDTH + threadIdx.x,
destY = dest / SharedDim_x, destX = dest % SharedDim_x,
srcY = blockIdx.y * TILE_HEIGHT + destY - SOBEL_OP_RADIUS,
srcX = blockIdx.x * TILE_WIDTH + destX - SOBEL_OP_RADIUS,
src = (srcY * width + srcX);
if (srcY >= 0 && srcY < height && srcX >= 0 && srcX < width)
N_ds[destY][destX] = inputImgData[src];
else
N_ds[destY][destX] = 0;
for (int iter = 1; iter <= (SharedDim_x * SharedDim_y) / (TILE_WIDTH * TILE_HEIGHT); iter++)
{
// other batch loading
dest = threadIdx.y * TILE_WIDTH + threadIdx.x + iter * (TILE_WIDTH * TILE_HEIGHT);
destY = dest / SharedDim_x, destX = dest % SharedDim_x;
srcY = blockIdx.y * TILE_HEIGHT + destY - SOBEL_OP_RADIUS;
srcX = blockIdx.x * TILE_WIDTH + destX - SOBEL_OP_RADIUS;
src = (srcY * width + srcX);
if (destY < SharedDim_y && destX < SharedDim_x)
{
if (srcY >= 0 && srcY < height && srcX >= 0 && srcX < width)
N_ds[destY][destX] = inputImgData[src];
else
N_ds[destY][destX] = 0;
}
}
__syncthreads();
int temp = 0;
int y, x;
for (y = 0; y < SOBEL_OP_DIM; y++)
for (x = 0; x < SOBEL_OP_DIM; x++)
temp += N_ds[threadIdx.y + y][threadIdx.x + x] * constMask[y * SOBEL_OP_DIM + x];
y = blockIdx.y * TILE_HEIGHT + threadIdx.y;
x = blockIdx.x * TILE_WIDTH + threadIdx.x;
if (y < height && x < width) {
gradientImgData[y * width + x] = (uint8_t)abs(temp);
}
__syncthreads();
}
according to nvprof shared memory implementations is slower:
Type Time(%) Time Calls Avg Min Max Name
GPU activities: 53.27% 387.52us 2 193.76us 190.70us 196.82us [CUDA memcpy DtoH]
24.28% 176.62us 2 88.311us 608ns 176.01us [CUDA memcpy HtoD]
11.56% 84.102us 1 84.102us 84.102us 84.102us tiledGradientConvolution(unsigned char*, unsigned char*, int, int)
10.90% 79.270us 1 79.270us 79.270us 79.270us constGradientConvolution(unsigned char*, unsigned char*, int, int)
this is the configuration kernel:
#define SOBEL_OP_DIM 3
#define SOBEL_OP_RADIUS (SOBEL_OP_DIM / 2)
// tile dimension
#define TILE_WIDTH 16
#define TILE_HEIGHT 16
// Allocate masks in constant memory
__constant__ int constMask[SOBEL_OP_DIM * SOBEL_OP_DIM];
// Shared Memory Elements needed to be loaded as per Mask Size
#define SharedDim_x (TILE_WIDTH + SOBEL_OP_DIM - 1)
#define SharedDim_y (TILE_HEIGHT + SOBEL_OP_DIM - 1)
// in main code//
dim3 dimBlock(TILE_WIDTH, TILE_HEIGHT);
dim3 dimGrid((test.w + TILE_WIDTH - 1) / TILE_WIDTH, (test.h + TILE_HEIGHT - 1) / TILE_HEIGHT);
I expect shared memory to be faster, but i can figure out what cause conflincts in loading from global memory.
Any help would be appreciated. Thanks you in advance.