glsl - glsl 将 mat2x4 重新解释为 mat4x2

Question

我试图在计算中实现 8x8 dct 方法。当前的实现如下所示：

void dct_8(in mat2x4 data, out vec4 res0246, out vec4 res1357) {
    res0246 = dct8_matrix_even*(data[0] + data[1].wzyx);
    res1357 = dct8_matrix_odd*(data[0] - data[1].wzyx);
}

void dct_8x8(inout float data[8][8]){
    float temp_buffer[8][8];
    // horizontal dct
    for(int i = 0; i<8; ++i){
        // compiler please optimize....
        mat2x4 temp;
        temp[0].x = data[i][0];
        temp[0].y = data[i][1];
        temp[0].z = data[i][2];
        temp[0].w = data[i][3];
        temp[1].x = data[i][4];
        temp[1].y = data[i][5];
        temp[1].z = data[i][6];
        temp[1].w = data[i][7];
        vec4 res0246;
        vec4 res1357;
        dct_8(temp, res0246, res1357);
        temp_buffer[i][0] = res0246.x;
        temp_buffer[i][1] = res1357.x;
        temp_buffer[i][2] = res0246.y;
        temp_buffer[i][3] = res1357.y;
        temp_buffer[i][4] = res0246.z;
        temp_buffer[i][5] = res1357.z;
        temp_buffer[i][6] = res0246.w;
        temp_buffer[i][7] = res1357.w;
    }
    // vertical dct
    for(int i = 0; i<8; ++i){
        // compiler please optimize....
        mat2x4 temp;
        temp[0].x = temp_buffer[0][i];
        temp[0].y = temp_buffer[1][i];
        temp[0].z = temp_buffer[2][i];
        temp[0].w = temp_buffer[3][i];
        temp[1].x = temp_buffer[4][i];
        temp[1].y = temp_buffer[5][i];
        temp[1].z = temp_buffer[6][i];
        temp[1].w = temp_buffer[7][i];
        vec4 res0246;
        vec4 res1357;
        dct_8(temp, res0246, res1357);
        data[0][i] = res0246.x;
        data[1][i] = res1357.x;
        data[2][i] = res0246.y;
        data[3][i] = res1357.y;
        data[4][i] = res0246.z;
        data[5][i] = res1357.z;
        data[6][i] = res0246.w;
        data[7][i] = res1357.w;
    }
}

水平存储到缓冲区操作的混合可以简化为返回 mat2x4，然后将该 mat2x4 解释为 mat4x2，然后转置它（然后将项目存储在连续的 8 个浮点数中）。

编译器是否能够对此进行优化？如果没有，是否可以显式编写此 swizzle 操作？

（如果它可以是显式的，这个实现会更好吗？）

void transpose_8x8(inout mat4x4 data[2][2])
{
    mat4x4 temp;
    data[0][0] = transpose(data[0][0]);
    temp = transpose(data[0][1]);
    data[0][1] = transpose(data[1][0]);
    data[1][0] = temp;
    data[1][1] = transpose(data[1][1]);
}

void dct8x8_vertical(in mat4x4 data[2][2], out mat4x4 result[2][2]){
    for (int i = 0; i < 2; ++i) {
        for(int j = 0; j<8; ++j){
            // compiler please optimize....
            mat2x4 temp = mat2x4(data[j/4][0][j%4], data[j/4][1][j%4]);
            mat2x4 res;
            dct_8(temp, res);
            result[j/4][0][j%4] = res[0];
            result[j/4][1][j%4] = res[1];
        }
    }
}

void dct_8x8(inout mat4x4 data[2][2])
{
    mat4x4 temp_buffer[2][2];
    // vertical dct
    dct8x8_vertical(data, temp_buffer);
    // horizontal dct
    transpose_8x8(temp_buffer);
    dct8x8_vertical(temp_buffer, data);
    transpose_8x8(data);
}

标签： glslgpgpudct