Got first results with simple watermarking in camera firmware, see the seminar presentation (May 28, Univ. Salzburg).

Main problem is speed currently: just unpacking 9 MB of sensor data takes over 2 sec. The sensor data is stored 10 bits/pixel (packed).

p_row_buf = p_out_row_buf = (uint16 *) &buf_pos[PIXTOBYTES(RAW_LEFT_MAR$
        out_bit_buf = bit_buf = 0;

        for (c = 0; c < IMG_WIDTH - 8; c += 8) {
            bit_buf = *p_row_buf++;
            out_bit_buf = bit_buf >> 6;

            bit_buf = (bit_buf << 16) + *p_row_buf++;
            pixel = bit_buf >> 12 & 0x3ff; // process pixel
            WATERMARK_PIXEL(pixel);
            out_bit_buf = (out_bit_buf << 10) + pixel;
            *p_out_row_buf++ = out_bit_buf >> 4;

            out_bit_buf = (out_bit_buf << 10) + (bit_buf >> 2 & 0x3ff);
	    //... and so on, repeats after 8 pixels
        }

The ARM assembly language can handle shifts very efficiently; a mov instrution can have a third operand specifying an implicit shift! The following GCC-generated assembler code corresponds to the first statements above. Note the asr and asl -- shift right, shift left.

        ldrh    r3, [sl], #2
        mov     r6, r7, asr #6
        ldr     r5, [r2, #12]
        add     r7, r3, r7, asl #16
        tst     r9, r1
        mov     r3, r7, asr #12
        mov     r4, r3, asl #22

posted at: 21:30 | path: /projects/CHDK | permanent link