avfilter/af_afir: implement non-uniform partitioned convolution

Using multiple frequency delay lines.

doc/filters.texi

@@ -1231,14 +1231,14 @@ Set video stream size. This option is used only when @var{response} is enabled.
 Set video stream frame rate. This option is used only when @var{response} is enabled.
 
 @item minp
-Set minimal partition size used for convolution. Default is @var{16}.
+Set minimal partition size used for convolution. Default is @var{8192}.
 Allowed range is from @var{16} to @var{32768}.
 Lower values decreases latency at cost of higher CPU usage.
 
 @item maxp
 Set maximal partition size used for convolution. Default is @var{8192}.
 Allowed range is from @var{16} to @var{32768}.
-Lower values decreases latency at cost of higher CPU usage.
+Lower values may increase CPU usage.
 @end table
 
 @subsection Examples

libavfilter/af_afir.h

@@ -37,14 +37,18 @@ typedef struct AudioFIRSegment {
     int block_size;
     int fft_length;
     int coeff_size;
-    int segment_size;
+    int input_size;
+    int input_offset;
 
+    int *output_offset;
     int *part_index;
 
     AVFrame *sum;
     AVFrame *block;
     AVFrame *buffer;
     AVFrame *coeff;
+    AVFrame *input;
+    AVFrame *output;
 
     RDFTContext **rdft, **irdft;
 } AudioFIRSegment;
@@ -80,6 +84,7 @@ typedef struct AudioFIRContext {
 
     AVFrame *in[2];
     AVFrame *video;
+    int min_part_size;
     int64_t pts;
 
     AVFloatDSPContext *fdsp;