[FFmpeg-devel] [PATCH 2/3] Indeo 5 decoder: common DSP functions
Maxim
max_pole
Sat Jan 9 01:01:34 CET 2010
Michael Niedermayer schrieb:
> On Sun, Jan 03, 2010 at 12:56:36PM +0200, Kostya wrote:
> [...]
>
>> void ff_ivi_recompose53(const IVIPlaneDesc *plane, uint8_t *dst,
>> const int dst_pitch, const int num_bands)
>> {
>> int x, y, indx;
>> int32_t p0, p1, p2, p3, tmp0, tmp1, tmp2;
>> int32_t b0_1, b0_2, b1_1, b1_2, b1_3, b2_1, b2_2, b2_3, b2_4, b2_5, b2_6;
>> int32_t b3_1, b3_2, b3_3, b3_4, b3_5, b3_6, b3_7, b3_8, b3_9;
>> uint32_t pitch, back_pitch;
>> const IDWTELEM *b0_ptr, *b1_ptr, *b2_ptr, *b3_ptr;
>>
>> /* all bands should have the same pitch */
>> pitch = plane->bands[0].pitch;
>>
>> /* pixels at the position "y-1" will be set to pixels at the "y" for the 1st iteration */
>> back_pitch = 0;
>>
>> /* get pointers to the wavelet bands */
>> b0_ptr = plane->bands[0].buf;
>> b1_ptr = plane->bands[1].buf;
>> b2_ptr = plane->bands[2].buf;
>> b3_ptr = plane->bands[3].buf;
>>
>> for (y = 0; y < plane->height; y += 2) {
>> /* load storage variables with values */
>> if (num_bands > 0) {
>> b0_1 = b0_ptr[0];
>> b0_2 = b0_ptr[pitch];
>> }
>>
>> if (num_bands > 1) {
>> b1_1 = b1_ptr[back_pitch];
>> b1_2 = b1_ptr[0];
>> b1_3 = b1_1 - b1_2*6 + b1_ptr[pitch];
>> }
>>
>> if (num_bands > 2) {
>> b2_2 = b2_ptr[0]; // b2[x, y ]
>> b2_3 = b2_2; // b2[x+1,y ] = b2[x,y]
>> b2_5 = b2_ptr[pitch]; // b2[x ,y+1]
>> b2_6 = b2_5; // b2[x+1,y+1] = b2[x,y+1]
>> }
>>
>> if (num_bands > 3) {
>> b3_2 = b3_ptr[back_pitch]; // b3[x ,y-1]
>> b3_3 = b3_2; // b3[x+1,y-1] = b3[x ,y-1]
>> b3_5 = b3_ptr[0]; // b3[x ,y ]
>> b3_6 = b3_5; // b3[x+1,y ] = b3[x ,y ]
>> b3_8 = b3_2 - b3_5*6 + b3_ptr[pitch];
>> b3_9 = b3_8;
>> }
>>
>> for (x = 0, indx = 0; x < plane->width; x+=2, indx++) {
>> /* some values calculated in the previous iterations can */
>> /* be reused in the next ones, so do appropriate copying */
>> b2_1 = b2_2; // b2[x-1,y ] = b2[x, y ]
>> b2_2 = b2_3; // b2[x ,y ] = b2[x+1,y ]
>> b2_4 = b2_5; // b2[x-1,y+1] = b2[x ,y+1]
>> b2_5 = b2_6; // b2[x ,y+1] = b2[x+1,y+1]
>> b3_1 = b3_2; // b3[x-1,y-1] = b3[x ,y-1]
>> b3_2 = b3_3; // b3[x ,y-1] = b3[x+1,y-1]
>> b3_4 = b3_5; // b3[x-1,y ] = b3[x ,y ]
>> b3_5 = b3_6; // b3[x ,y ] = b3[x+1,y ]
>> b3_7 = b3_8; // vert_HPF(x-1)
>> b3_8 = b3_9; // vert_HPF(x )
>>
>> p0 = p1 = p2 = p3 = 0;
>>
>> /* process the LL-band by applying LPF both vertically and horizontally */
>> if (num_bands > 0) {
>> tmp0 = b0_1;
>> tmp2 = b0_2;
>> b0_1 = b0_ptr[indx+1];
>> b0_2 = b0_ptr[pitch+indx+1];
>> tmp1 = tmp0 + b0_1;
>>
>> p0 = tmp0 << 4;
>> p1 = tmp1 << 3;
>> p2 = (tmp0 + tmp2) << 3;
>> p3 = (tmp1 + tmp2 + b0_2) << 2;
>> }
>>
>> /* process the HL-band by applying HPF vertically and LPF horizontally */
>> if (num_bands > 1) {
>> tmp0 = b1_2;
>> tmp1 = b1_1;
>> b1_2 = b1_ptr[indx+1];
>> b1_1 = b1_ptr[back_pitch+indx+1];
>>
>> tmp2 = tmp1 - tmp0*6 + b1_3;
>> b1_3 = b1_1 - b1_2*6 + b1_ptr[pitch+indx+1];
>>
>> p0 += (tmp0 + tmp1) << 3;
>> p1 += (tmp0 + tmp1 + b1_1 + b1_2) << 2;
>> p2 += tmp2 << 2;
>> p3 += (tmp2 + b1_3) << 1;
>> }
>>
>> /* process the LH-band by applying LPF vertically and HPF horizontally */
>> if (num_bands > 2) {
>> b2_3 = b2_ptr[indx+1];
>> b2_6 = b2_ptr[pitch+indx+1];
>>
>> tmp0 = b2_1 + b2_2;
>> tmp1 = b2_1 - b2_2*6 + b2_3;
>>
>> p0 += tmp0 << 3;
>> p1 += tmp1 << 2;
>> p2 += (tmp0 + b2_4 + b2_5) << 2;
>> p3 += (tmp1 + b2_4 - b2_5*6 + b2_6) << 1;
>> }
>>
>> /* process the HH-band by applying HPF both vertically and horizontally */
>> if (num_bands > 3) {
>> b3_6 = b3_ptr[indx+1]; // b3[x+1,y ]
>> b3_3 = b3_ptr[back_pitch+indx+1]; // b3[x+1,y-1]
>>
>> tmp0 = b3_1 + b3_4;
>> tmp1 = b3_2 + b3_5;
>> tmp2 = b3_3 + b3_6;
>>
>> b3_9 = b3_3 - b3_6*6 + b3_ptr[pitch+indx+1];
>>
>> p0 += (tmp0 + tmp1) << 2;
>> p1 += (tmp0 - tmp1*6 + tmp2) << 1;
>> p2 += (b3_7 + b3_8) << 1;
>> p3 += b3_7 - b3_8*6 + b3_9;
>> }
>>
>> /* output four pixels */
>> dst[x] = av_clip_uint8((p0 >> 6) + 128);
>> dst[x+1] = av_clip_uint8((p1 >> 6) + 128);
>> dst[dst_pitch+x] = av_clip_uint8((p2 >> 6) + 128);
>> dst[dst_pitch+x+1] = av_clip_uint8((p3 >> 6) + 128);
>> }// for x
>>
>> dst += dst_pitch << 1;
>>
>> back_pitch = -pitch;
>>
>> b0_ptr += pitch;
>> b1_ptr += pitch;
>> b2_ptr += pitch;
>> b3_ptr += pitch;
>> }
>> }
>>
>
> is this mess faster than some more readable variant?
>
It resembles the original code and was include for testing purposes!
Xanim/Mac indeo5 binaries contain really buggy code for this function :-(
My intention was to reuse the snow's compose53i code at the end! So I
suggest to split that code out of the snow's source and place it into a
separate file. BTW JPEG2000 uses the same wavelet. Once thing required
is a scale step for indeo5 HPF coefficients. Maybe that could be
incorporated into the code...
Regards
maxim
More information about the ffmpeg-devel
mailing list