E.x - 1 < srcBoundary.x) {
            mask &= 0xff00ffffu;
            v[2] = textureLoadGeneral(src_tex, coordE - vec3u(1, 0, 0), params.mipLevel).r;
        }

        if (coordE.x >= 2) {
            if (coordE.x - 2 < srcBoundary.x) {
                mask &= 0xffff00ffu;
                v[1] = textureLoadGeneral(src_tex, coordE - vec3u(2, 0, 0), params.mipLevel).r;
            }

            if (coordE.x >= 3) {
                if (coordE.x - 3 < srcBoundary.x) {
                    mask &= 0xffffff00u;
                    v[0] = textureLoadGeneral(src_tex, coordE - vec3u(3, 0, 0), params.mipLevel).r;
                }
            }
        }
    }
}

if (coordE.x < srcBoundary.x) {
    mask &= 0x00ffffffu;
    v[3] = textureLoadGeneral(src_tex, coordE, params.mipLevel).r;
} else {
    // coordE is not used
    // dstBuf value is used for later part.
    // end of buffer (last thread) / end of non-compact row + x boundary check
    if (coordE.x - 2 < srcBoundary.x) {
        mask &= 0xffff00ffu;
        v[1] = textureLoadGeneral(src_tex, coordE - vec3u(2, 0, 0), params.mipLevel).r;
        if (coordE.x - 1 < srcBoundary.x) {
            mask &= 0xff00ffffu;
            v[2] = textureLoadGeneral(src_tex, coordE - vec3u(1, 0, 0), params.mipLevel).r;
        }
    }
}

if (readDstBufAtStart || readDstBufAtEnd) {
    let original: u32 = dst_buf[dstOffset];
    result = (original & mask) | (encodeVectorInU32General(v) & ~mask);
} else {
    var coord1: vec3u;
    var coord2: vec3u;
    if (coordS.x < coordE.x) {
        // middle of row
        coord1 = coordE - vec3u(2, 0, 0);
        coord2 = coordE - vec3u(1, 0, 0);
    } else {
        // start of row
        switch params.shift {
            case 0: {
                coord1 = coordS + vec3u(1, 0, 0);
                coord2 = coordS + vec3u(2, 0, 0);
            }
            case 1: {
                coord1 = coordE - vec3u(2, 0, 0);
                coord2 = coordE - vec3u(1, 0, 0);
            }
            case 2: {
                coord1 = coordS + vec3u(1, 0, 0);
                coord2 = coordE - vec3u(1, 0, 0);
            }
            case 3: {
                coord1 = coordS + vec3u(1, 0, 0);
                coord2 = coordS + vec3u(2, 0, 0);
            }
            default: {
                return; // unreachable when shift == 0
            }
        }
    }

    if (coord1.x < srcBoundary.x) {
        mask &= 0xffff00ffu;
        v[1] = textureLoadGeneral(src_tex, coord1, params.mipLevel).r;
    }
    if (coord2.x < srcBoundary.x) {
        mask &= 0xff00ffffu;
        v[2] = textureLoadGeneral(src_tex, coord2, params.mipLevel).r;
    }

    let readDstBufAtMid: bool = (params.srcExtent.x + params.shift > params.bytesPerRow)
        && (params.srcExtent.x < params.bytesPerRow);
    if (readDstBufAtMid && id.x == 0) {
        let original: u32 = dst_buf[dstOffset];
        result = (original & mask) | (encodeVectorInU32General(v) & ~mask);
    } else {
        result = encodeVectorInU32General(v);
    }
}