From: Sebastian Dröge Date: Thu, 3 Dec 2009 16:27:13 +0000 (+0100) Subject: audiofxbasefirfilter: FFT convolution implementation X-Git-Tag: 1.19.3~509^2~9363 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=02960383c1b1b65ead06719fc60db98410afe586;p=platform%2Fupstream%2Fgstreamer.git audiofxbasefirfilter: FFT convolution implementation This provides a great speedup, especially the relationship between kernel length and processing size is now logarithmic instead of linear. Below a kernel size of 32 it's a bit slower, afterwards it's much faster: 17 0.788000 -> 0.950000 33 1.208000 -> 1.146000 65 2.166000 -> 1.146000 ... 4097 107.444000 -> 1.508000 For sizes smaller 32 the normal time-domain convolution is chosen, for larger sizes the FFT convolution is automatically used. Fixes bug #594381. --- diff --git a/gst/audiofx/Makefile.am b/gst/audiofx/Makefile.am index 22f5fd0..bd66963 100644 --- a/gst/audiofx/Makefile.am +++ b/gst/audiofx/Makefile.am @@ -29,6 +29,7 @@ libgstaudiofx_la_LIBADD = $(GST_LIBS) \ $(GST_CONTROLLER_LIBS) \ $(GST_PLUGINS_BASE_LIBS) \ -lgstaudio-$(GST_MAJORMINOR) \ + -lgstfft-$(GST_MAJORMINOR) \ $(LIBM) libgstaudiofx_la_LDFLAGS = $(GST_PLUGIN_LDFLAGS) libgstaudiofx_la_LIBTOOLFLAGS = --tag=disable-static diff --git a/gst/audiofx/audiofxbasefirfilter.c b/gst/audiofx/audiofxbasefirfilter.c index c66bd26..c75f98d 100644 --- a/gst/audiofx/audiofxbasefirfilter.c +++ b/gst/audiofx/audiofxbasefirfilter.c @@ -59,6 +59,9 @@ GST_DEBUG_CATEGORY_STATIC (GST_CAT_DEFAULT); GST_DEBUG_CATEGORY_INIT (gst_audio_fx_base_fir_filter_debug, "audiofxbasefirfilter", 0, \ "FIR filter base class"); +/* Switch from time-domain to FFT convolution for kernels >= this */ +#define FFT_THRESHOLD 32 + GST_BOILERPLATE_FULL (GstAudioFXBaseFIRFilter, gst_audio_fx_base_fir_filter, GstAudioFilter, GST_TYPE_AUDIO_FILTER, DEBUG_INIT); @@ -68,6 +71,9 @@ static gboolean gst_audio_fx_base_fir_filter_start (GstBaseTransform * base); static gboolean gst_audio_fx_base_fir_filter_stop (GstBaseTransform * base); static gboolean gst_audio_fx_base_fir_filter_event (GstBaseTransform * base, GstEvent * event); +static gboolean gst_audio_fx_base_fir_filter_transform_size (GstBaseTransform * + base, GstPadDirection direction, GstCaps * caps, guint size, + GstCaps * othercaps, guint * othersize); static gboolean gst_audio_fx_base_fir_filter_setup (GstAudioFilter * base, GstRingBufferSpec * format); @@ -83,16 +89,23 @@ gst_audio_fx_base_fir_filter_dispose (GObject * object) { GstAudioFXBaseFIRFilter *self = GST_AUDIO_FX_BASE_FIR_FILTER (object); - if (self->buffer) { - g_free (self->buffer); - self->buffer = NULL; - self->buffer_length = 0; - } + g_free (self->buffer); + self->buffer = NULL; + self->buffer_length = 0; - if (self->kernel) { - g_free (self->kernel); - self->kernel = NULL; - } + g_free (self->kernel); + self->kernel = NULL; + + gst_fft_f64_free (self->fft); + self->fft = NULL; + gst_fft_f64_free (self->ifft); + self->ifft = NULL; + + g_free (self->frequency_response); + self->frequency_response = NULL; + + g_free (self->fft_buffer); + self->fft_buffer = NULL; G_OBJECT_CLASS (parent_class)->dispose (object); } @@ -122,6 +135,8 @@ gst_audio_fx_base_fir_filter_class_init (GstAudioFXBaseFIRFilterClass * klass) trans_class->start = GST_DEBUG_FUNCPTR (gst_audio_fx_base_fir_filter_start); trans_class->stop = GST_DEBUG_FUNCPTR (gst_audio_fx_base_fir_filter_stop); trans_class->event = GST_DEBUG_FUNCPTR (gst_audio_fx_base_fir_filter_event); + trans_class->transform_size = + GST_DEBUG_FUNCPTR (gst_audio_fx_base_fir_filter_transform_size); filter_class->setup = GST_DEBUG_FUNCPTR (gst_audio_fx_base_fir_filter_setup); } @@ -144,6 +159,176 @@ gst_audio_fx_base_fir_filter_init (GstAudioFXBaseFIRFilter * self, gst_audio_fx_base_fir_filter_query_type); } +/* This implements FFT convolution and uses the overlap-save algorithm. + * See http://cnx.org/content/m12022/latest/ or your favorite + * digital signal processing book for details. + * + * In every pass the following is calculated: + * + * y = IFFT (FFT(x) * FFT(h)) + * + * where y is the output in the time domain, x the + * input and h the filter kernel. * is the multiplication + * of complex numbers. + * + * Due to the circular convolution theorem this + * gives in the time domain: + * + * y[t] = \sum_{u=0}^{M-1} x[t - u] * h[u] + * + * where y is the output, M is the kernel length, + * x the periodically extended[0] input and h the + * filter kernel. + * + * ([0] Periodically extended means: ) + * ( x[t] = x[t+kN] \forall k \in Z ) + * ( where N is the length of x ) + * + * This means: + * - Obviously x and h need to be of the same size for the FFT + * - The first M-1 output values are useless because they're + * built from 1 up to M-1 values from the end of the input + * (circular convolusion!). + * - The last M-1 input values are only used for 1 up to M-1 + * output values, i.e. they need to be used again in the + * next pass for the first M-1 input values. + * + * => The first pass needs M-1 zeroes at the beginning of the + * input and the last M-1 input values of every pass need to + * be used as the first M-1 input values of the next pass. + * + * => x must be larger than h to give a useful number of output + * samples and h needs to be padded by zeroes at the end to give + * it virtually the same size as x (by M we denote the number of + * non-padding samples of h). If len(x)==len(h)==M only 1 output + * sample would be calculated per pass, len(x)==2*len(h) would + * give M+1 output samples, etc. Usually a factor between 4 and 8 + * gives a low number of operations per output samples (see website + * given above). + * + * Overall this gives a runtime complexity per sample of + * + * ( N log N ) + * O ( --------- ) compared to O (M) for the direct calculation. + * ( N - M + 1 ) + */ +#define DEFINE_FFT_PROCESS_FUNC(width,ctype) \ +static guint \ +process_fft_##width (GstAudioFXBaseFIRFilter * self, const g##ctype * src, \ + g##ctype * dst, guint input_samples) \ +{ \ + gint channels = GST_AUDIO_FILTER_CAST (self)->format.channels; \ + gint i, j; \ + guint pass; \ + guint kernel_length = self->kernel_length; \ + guint block_length = self->block_length; \ + guint buffer_length = self->buffer_length; \ + guint real_buffer_length = buffer_length + kernel_length - 1; \ + guint buffer_fill = self->buffer_fill; \ + GstFFTF64 *fft = self->fft; \ + GstFFTF64 *ifft = self->ifft; \ + GstFFTF64Complex *frequency_response = self->frequency_response; \ + GstFFTF64Complex *fft_buffer = self->fft_buffer; \ + guint frequency_response_length = self->frequency_response_length; \ + gdouble *buffer = self->buffer; \ + guint generated = 0; \ + gdouble re, im; \ + \ + input_samples /= channels; \ + \ + if (!fft_buffer) \ + self->fft_buffer = fft_buffer = \ + g_new (GstFFTF64Complex, frequency_response_length); \ + \ + /* Buffer contains the time domain samples of input data for one chunk \ + * plus some more space for the inverse FFT below. \ + * \ + * The samples are put at offset kernel_length, the inverse FFT \ + * overwrites everthing from offset 0 to length-kernel_length+1, keeping \ + * the last kernel_length-1 samples for copying to the next processing \ + * step. \ + */ \ + if (!buffer) { \ + self->buffer_length = buffer_length = block_length; \ + real_buffer_length = buffer_length + kernel_length - 1; \ + \ + self->buffer = buffer = g_new0 (gdouble, real_buffer_length * channels); \ + \ + /* Beginning has kernel_length-1 zeroes at the beginning */ \ + self->buffer_fill = buffer_fill = kernel_length - 1; \ + } \ + \ + while (input_samples) { \ + pass = MIN (buffer_length - buffer_fill, input_samples); \ + \ + /* Deinterleave channels */ \ + for (i = 0; i < pass; i++) { \ + for (j = 0; j < channels; j++) { \ + buffer[real_buffer_length * j + buffer_fill + kernel_length - 1 + i] = \ + src[i * channels + j]; \ + } \ + } \ + buffer_fill += pass; \ + src += channels * pass; \ + input_samples -= channels * pass; \ + \ + /* If we don't have a complete buffer go out */ \ + if (buffer_fill < buffer_length) \ + break; \ + \ + for (j = 0; j < channels; j++) { \ + /* Calculate FFT of input block */ \ + gst_fft_f64_fft (fft, \ + buffer + real_buffer_length * j + kernel_length - 1, fft_buffer); \ + \ + /* Complex multiplication of input and filter spectrum */ \ + for (i = 0; i < frequency_response_length; i++) { \ + re = fft_buffer[i].r; \ + im = fft_buffer[i].i; \ + \ + fft_buffer[i].r = \ + re * frequency_response[i].r - \ + im * frequency_response[i].i; \ + fft_buffer[i].i = \ + re * frequency_response[i].i + \ + im * frequency_response[i].r; \ + } \ + \ + /* Calculate inverse FFT of the result */ \ + gst_fft_f64_inverse_fft (ifft, fft_buffer, \ + buffer + real_buffer_length * j); \ + \ + /* Copy all except the first kernel_length-1 samples to the output */ \ + for (i = 0; i < buffer_length - kernel_length + 1; i++) { \ + dst[i * channels + j] = \ + buffer[real_buffer_length * j + kernel_length - 1 + i]; \ + } \ + \ + /* Copy the last kernel_length-1 samples to the beginning for the next block */ \ + for (i = 0; i < kernel_length - 1; i++) { \ + buffer[real_buffer_length * j + kernel_length - 1 + i] = \ + buffer[real_buffer_length * j + buffer_length + i]; \ + } \ + } \ + \ + generated += buffer_length - kernel_length + 1; \ + dst += channels * (buffer_length - kernel_length + 1); \ + \ + /* The the first kernel_length-1 samples are there already */ \ + buffer_fill = kernel_length - 1; \ + } \ + \ + /* Write back cached buffer_fill value */ \ + self->buffer_fill = buffer_fill; \ + \ + return generated; \ +} + +DEFINE_FFT_PROCESS_FUNC (32, float); +DEFINE_FFT_PROCESS_FUNC (64, double); + +#undef DEFINE_FFT_PROCESS_FUNC + /* * The code below calculates the linear convolution: * @@ -231,7 +416,6 @@ gst_audio_fx_base_fir_filter_push_residue (GstAudioFXBaseFIRFilter * self) gint channels = GST_AUDIO_FILTER_CAST (self)->format.channels; gint width = GST_AUDIO_FILTER_CAST (self)->format.width / 8; guint outsize, outsamples; - gint64 diffsize, diffsamples; guint8 *in, *out; if (channels == 0 || rate == 0 || self->nsamples_in == 0) { @@ -252,39 +436,66 @@ gst_audio_fx_base_fir_filter_push_residue (GstAudioFXBaseFIRFilter * self) } outsize = outsamples * channels * width; - /* Process the difference between latency and residue length samples - * to start at the actual data instead of starting at the zeros before - * when we only got one buffer smaller than latency */ - - /* FIXME: still time domain convolution specific */ - diffsamples = - ((gint64) self->latency) - ((gint64) self->buffer_fill) / channels; - if (diffsamples > 0) { - diffsize = diffsamples * channels * width; - in = g_new0 (guint8, diffsize); - out = g_new0 (guint8, diffsize); - self->nsamples_out += self->process (self, in, out, diffsamples * channels); + if (!self->fft) { + gint64 diffsize, diffsamples; + + /* Process the difference between latency and residue length samples + * to start at the actual data instead of starting at the zeros before + * when we only got one buffer smaller than latency */ + diffsamples = + ((gint64) self->latency) - ((gint64) self->buffer_fill) / channels; + if (diffsamples > 0) { + diffsize = diffsamples * channels * width; + in = g_new0 (guint8, diffsize); + out = g_new0 (guint8, diffsize); + self->nsamples_out += + self->process (self, in, out, diffsamples * channels); + g_free (in); + g_free (out); + } + + res = gst_pad_alloc_buffer (GST_BASE_TRANSFORM_CAST (self)->srcpad, + GST_BUFFER_OFFSET_NONE, outsize, + GST_PAD_CAPS (GST_BASE_TRANSFORM_CAST (self)->srcpad), &outbuf); + + if (G_UNLIKELY (res != GST_FLOW_OK)) { + GST_WARNING_OBJECT (self, "failed allocating buffer of %d bytes", + outsize); + self->buffer_fill = 0; + return; + } + + /* Convolve the residue with zeros to get the actual remaining data */ + in = g_new0 (guint8, outsize); + self->nsamples_out += + self->process (self, in, GST_BUFFER_DATA (outbuf), + outsamples * channels); g_free (in); - g_free (out); - } + } else { + guint gensamples = 0; + guint8 *data; - res = gst_pad_alloc_buffer (GST_BASE_TRANSFORM_CAST (self)->srcpad, - GST_BUFFER_OFFSET_NONE, outsize, - GST_PAD_CAPS (GST_BASE_TRANSFORM_CAST (self)->srcpad), &outbuf); + outbuf = gst_buffer_new_and_alloc (outsize); + data = GST_BUFFER_DATA (outbuf); - if (G_UNLIKELY (res != GST_FLOW_OK)) { - GST_WARNING_OBJECT (self, "failed allocating buffer of %d bytes", outsize); - self->buffer_fill = 0; - return; - } + while (gensamples < outsamples) { + guint step_insamples = + (self->block_length - self->buffer_fill) * channels; + guint8 *zeroes = g_new0 (guint8, step_insamples * width); + guint8 *out = g_new (guint8, self->block_length * channels * width); + guint step_gensamples; - /* Convolve the residue with zeros to get the actual remaining data */ - in = g_new0 (guint8, outsize); - self->nsamples_out += - self->process (self, in, GST_BUFFER_DATA (outbuf), outsamples * channels); - g_free (in); + step_gensamples = self->process (self, zeroes, out, step_insamples); + g_free (zeroes); - /* FIXME: time domain convolution specific */ + memcpy (data + gensamples * width, out, MIN (step_gensamples, + outsamples - gensamples) * width); + gensamples += MIN (step_gensamples, outsamples - gensamples); + + g_free (out); + } + self->nsamples_out += gensamples; + } /* Set timestamp, offset, etc from the values we * saved when processing the regular buffers */ @@ -343,18 +554,53 @@ gst_audio_fx_base_fir_filter_setup (GstAudioFilter * base, self->nsamples_in = 0; } - if (format->width == 32) + if (format->width == 32 && self->fft) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_fft_32; + else if (format->width == 64 && self->fft) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_fft_64; + else if (format->width == 32) self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_32; else if (format->width == 64) self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_64; - else - ret = FALSE; + ret = FALSE; return TRUE; } /* GstBaseTransform vmethod implementations */ +static gboolean +gst_audio_fx_base_fir_filter_transform_size (GstBaseTransform * base, + GstPadDirection direction, GstCaps * caps, guint size, GstCaps * othercaps, + guint * othersize) +{ + GstAudioFXBaseFIRFilter *self = GST_AUDIO_FX_BASE_FIR_FILTER (base); + guint blocklen; + GstStructure *s; + gint width, channels; + + if (!self->fft || direction == GST_PAD_SRC) { + *othersize = size; + return TRUE; + } + + s = gst_caps_get_structure (caps, 0); + if (!gst_structure_get_int (s, "width", &width) || + !gst_structure_get_int (s, "channels", &channels)) + return FALSE; + + width /= 8; + + size /= width * channels; + + blocklen = self->block_length - self->kernel_length + 1; + *othersize = ((size + blocklen - 1) / blocklen) * blocklen; + + *othersize *= width * channels; + + return TRUE; +} + static GstFlowReturn gst_audio_fx_base_fir_filter_transform (GstBaseTransform * base, GstBuffer * inbuf, GstBuffer * outbuf) @@ -512,9 +758,13 @@ gst_audio_fx_base_fir_filter_query (GstPad * pad, GstQuery * query) GST_TIME_FORMAT " max %" GST_TIME_FORMAT, GST_TIME_ARGS (min), GST_TIME_ARGS (max)); + if (self->fft) + latency = self->block_length - self->kernel_length + 1; + else + latency = self->latency; + /* add our own latency */ - latency = - gst_util_uint64_scale_round (self->latency, GST_SECOND, rate); + latency = gst_util_uint64_scale_round (latency, GST_SECOND, rate); GST_DEBUG_OBJECT (self, "Our latency: %" GST_TIME_FORMAT, GST_TIME_ARGS (latency)); @@ -576,6 +826,11 @@ void gst_audio_fx_base_fir_filter_set_kernel (GstAudioFXBaseFIRFilter * self, gdouble * kernel, guint kernel_length, guint64 latency) { + gdouble *kernel_tmp; + guint i; + gboolean latency_changed; + gint width; + g_return_if_fail (kernel != NULL); g_return_if_fail (self != NULL); @@ -589,16 +844,64 @@ gst_audio_fx_base_fir_filter_set_kernel (GstAudioFXBaseFIRFilter * self, self->buffer_fill = 0; } + latency_changed = (self->latency != latency + || (self->kernel_length < FFT_THRESHOLD && kernel_length >= FFT_THRESHOLD) + || (self->kernel_length >= FFT_THRESHOLD + && kernel_length < FFT_THRESHOLD)); + g_free (self->kernel); g_free (self->buffer); self->buffer = NULL; self->buffer_fill = 0; self->buffer_length = 0; + gst_fft_f64_free (self->fft); + self->fft = NULL; + gst_fft_f64_free (self->ifft); + self->ifft = NULL; + g_free (self->frequency_response); + self->frequency_response_length = 0; + g_free (self->fft_buffer); + self->fft_buffer = NULL; + self->kernel = kernel; self->kernel_length = kernel_length; - if (self->latency != latency) { + if (kernel_length >= FFT_THRESHOLD) { + /* We process 4 * kernel_length samples per pass in FFT mode */ + kernel_length = 4 * kernel_length; + kernel_length = gst_fft_next_fast_length (kernel_length); + self->block_length = kernel_length; + + kernel_tmp = g_new0 (gdouble, kernel_length); + memcpy (kernel_tmp, kernel, self->kernel_length * sizeof (gdouble)); + + self->fft = gst_fft_f64_new (kernel_length, FALSE); + self->ifft = gst_fft_f64_new (kernel_length, TRUE); + self->frequency_response_length = kernel_length / 2 + 1; + self->frequency_response = + g_new (GstFFTF64Complex, self->frequency_response_length); + gst_fft_f64_fft (self->fft, kernel_tmp, self->frequency_response); + g_free (kernel_tmp); + + /* Normalize to make sure IFFT(FFT(x)) == x */ + for (i = 0; i < self->frequency_response_length; i++) { + self->frequency_response[i].r /= kernel_length; + self->frequency_response[i].i /= kernel_length; + } + } + + width = GST_AUDIO_FILTER_CAST (self)->format.width; + if (width == 32 && self->fft) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_fft_32; + else if (width == 64 && self->fft) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_fft_64; + else if (width == 32) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_32; + else if (width == 64) + self->process = (GstAudioFXBaseFIRFilterProcessFunc) process_64; + + if (latency_changed) { self->latency = latency; gst_element_post_message (GST_ELEMENT (self), gst_message_new_latency (GST_OBJECT (self))); diff --git a/gst/audiofx/audiofxbasefirfilter.h b/gst/audiofx/audiofxbasefirfilter.h index aa03b1c..fd3c3bd 100644 --- a/gst/audiofx/audiofxbasefirfilter.h +++ b/gst/audiofx/audiofxbasefirfilter.h @@ -27,6 +27,7 @@ #include #include +#include G_BEGIN_DECLS @@ -54,17 +55,26 @@ typedef guint (*GstAudioFXBaseFIRFilterProcessFunc) (GstAudioFXBaseFIRFilter *, struct _GstAudioFXBaseFIRFilter { GstAudioFilter element; - /* < private > */ - GstAudioFXBaseFIRFilterProcessFunc process; - + /* properties */ gdouble *kernel; /* filter kernel -- time domain */ guint kernel_length; /* length of the filter kernel -- time domain */ + guint64 latency; /* pre-latency of the filter kernel */ + + /* < private > */ + GstAudioFXBaseFIRFilterProcessFunc process; + gdouble *buffer; /* buffer for storing samples of previous buffers */ guint buffer_fill; /* fill level of buffer */ - guint buffer_length; /* length of the buffer */ - - guint64 latency; + guint buffer_length; /* length of the buffer -- meaning depends on processing mode */ + + /* FFT convolution specific data */ + GstFFTF64 *fft; + GstFFTF64 *ifft; + GstFFTF64Complex *frequency_response; /* filter kernel -- frequency domain */ + guint frequency_response_length; /* length of filter kernel -- frequency domain */ + GstFFTF64Complex *fft_buffer; /* FFT buffer, has the length of the frequency response */ + guint block_length; /* Length of the processing blocks -- time domain */ GstClockTime start_ts; /* start timestamp after a discont */ guint64 start_off; /* start offset after a discont */