simple lossy procedure preceding lz4 compression

lossy_lz4_rgb.c

// simple lossy procedure preceding lz4 compression
// default: win_size = 64, max_loss = 2
// tested with lz4hc - LZ4HC_CLEVEL_DEFAULT

void lossy_lz4_rgb(uint8_t *data, int w, int h, int win_size, float max_loss)
{
	float max_loss2 = max_loss*max_loss;
	int y, x;

	for (y = 0; y < h; y++)
	for (x = 0; x < w; x++) {

		uint8_t *pixel;
		float r, g, b;
		int yy, xx, my, mx;

		// R2 sequence
		float r2 = 0.754877645 * x + 0.569840259 * y;
		float r2i = r2 - (int)r2;

		// triangulate and remap to max_loss
		float dither = r2i < 0.5 ? r2i = r2i * 2.0 : 2.0 - r2i * 2.0;
		dither = (dither * 2.0 - 1.0) * max_loss;

		// add dither (hurt compression but remove banding)
		pixel = data + (y * w + x) * 3;
		r = pixel[0] + dither;
		g = pixel[1] + dither;
		b = pixel[2] + dither;

		// window search starting from current pixel
		// lz4 max offset is 65535 bytes so it might not hold for large images
		my = M_MAX(0, y - win_size);
		mx = M_MAX(0, x - win_size);

		for (yy = y; yy >= my; yy--)
		for (xx = x; xx >= mx; xx--) {
			if (!(xx == x && yy == y)) {

				uint8_t *win_pixel = data + (yy * w + xx) * 3;
				float dr = win_pixel[0] - r;
				float dg = win_pixel[1] - g;
				float db = win_pixel[2] - b;
				float d = dr*dr*0.3 + dg*dg*0.5 + db*db*0.2;

				if (d < max_loss2) {
					pixel[0] = win_pixel[0];
					pixel[1] = win_pixel[1];
					pixel[2] = win_pixel[2];
					goto next_pixel;
				}
			}
		}

		next_pixel:;
	}
}