OrgACRN
/
sof
mirror of https://github.com/thesofproject/sof.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Audio: TDFB: Optimize pass-through MCPS for beamformer 

This patch adds use of dedicated pass-through functions to save
MCPS. The pass-through is used when ALSA beam control is to off.
The earlier pass-through mode was made with a minimum length set
channels filters for the FIR filter bank.

In TGL platform with 4 to 2 channels beamformer the patch
saves 18 MCPS, from 27 to 9 MCPS. With 2 to 2 channels beamformer
the patch saves 14 MCPS, from 16 to 2 MCPS.

Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>
This commit is contained in:
Seppo Ingalsuo 2024-05-20 20:00:53 +03:00 committed by Kai Vehmanen
parent e2dbe90358
commit 0b6cf0250e
1 changed files with 201 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

205
src/audio/tdfb/tdfb.c
View File

@ -79,6 +79,187 @@ static inline int set_func(struct processing_module *mod, enum sof_ipc_frame fmt
return 0;
}
/*
* Pass-through processing functions
*/
static void tdfb_pass_same_format(struct tdfb_comp_data *cd, struct input_stream_buffer *bsource,
struct output_stream_buffer *bsink, int frames)
{
struct audio_stream *source = bsource->data;
struct audio_stream *sink = bsink->data;
audio_stream_copy(source, 0, sink, 0, frames * audio_stream_get_channels(source));
}
#if CONFIG_FORMAT_S16LE
static void tdfb_pass_s16(struct tdfb_comp_data *cd, struct input_stream_buffer *bsource,
struct output_stream_buffer *bsink, int frames)
{
struct audio_stream *source = bsource->data;
struct audio_stream *sink = bsink->data;
int16_t *x = audio_stream_get_rptr(source);
int16_t *y = audio_stream_get_wptr(sink);
int16_t s0, s1;
int fmax;
int is, om;
int f, i, j, k;
const int in_nch = audio_stream_get_channels(source);
const int out_nch = audio_stream_get_channels(sink);
const int num_filters = cd->config->num_filters;
int remaining_frames = frames;
while (remaining_frames) {
fmax = audio_stream_frames_without_wrap(source, x);
f = MIN(remaining_frames, fmax);
fmax = audio_stream_frames_without_wrap(sink, y);
f = MIN(f, fmax);
for (j = 0; j < f; j += 2) {
for (i = 0; i < num_filters; i++) {
is = cd->input_channel_select[i];
om = cd->output_channel_mix[i];
s0 = x[is];
s1 = x[is + in_nch];
for (k = 0; k < out_nch; k++) {
if (om & 1) {
y[k] = s0;
y[k + out_nch] = s1;
}
om = om >> 1;
}
}
x += 2 * in_nch;
y += 2 * out_nch;
}
remaining_frames -= f;
x = audio_stream_wrap(source, x);
y = audio_stream_wrap(sink, y);
}
}
#endif
#if CONFIG_FORMAT_S24LE
static void tdfb_pass_s24(struct tdfb_comp_data *cd, struct input_stream_buffer *bsource,
struct output_stream_buffer *bsink, int frames)
{
struct audio_stream *source = bsource->data;
struct audio_stream *sink = bsink->data;
int32_t *x = audio_stream_get_rptr(source);
int32_t *y = audio_stream_get_wptr(sink);
int32_t s0, s1;
int fmax;
int is, om;
int f, i, j, k;
const int in_nch = audio_stream_get_channels(source);
const int out_nch = audio_stream_get_channels(sink);
const int num_filters = cd->config->num_filters;
int remaining_frames = frames;
while (remaining_frames) {
fmax = audio_stream_frames_without_wrap(source, x);
f = MIN(remaining_frames, fmax);
fmax = audio_stream_frames_without_wrap(sink, y);
f = MIN(f, fmax);
for (j = 0; j < f; j += 2) {
for (i = 0; i < num_filters; i++) {
is = cd->input_channel_select[i];
om = cd->output_channel_mix[i];
s0 = x[is];
s1 = x[is + in_nch];
for (k = 0; k < out_nch; k++) {
if (om & 1) {
y[k] = s0;
y[k + out_nch] = s1;
}
om = om >> 1;
}
}
x += 2 * in_nch;
y += 2 * out_nch;
}
remaining_frames -= f;
x = audio_stream_wrap(source, x);
y = audio_stream_wrap(sink, y);
}
}
#endif
#if CONFIG_FORMAT_S32LE
static void tdfb_pass_s32(struct tdfb_comp_data *cd, struct input_stream_buffer *bsource,
struct output_stream_buffer *bsink, int frames)
{
struct audio_stream *source = bsource->data;
struct audio_stream *sink = bsink->data;
int32_t *x = audio_stream_get_rptr(source);
int32_t *y = audio_stream_get_wptr(sink);
int32_t s0, s1;
int fmax;
int is, om;
int f, i, j, k;
const int in_nch = audio_stream_get_channels(source);
const int out_nch = audio_stream_get_channels(sink);
const int num_filters = cd->config->num_filters;
int remaining_frames = frames;
while (remaining_frames) {
fmax = audio_stream_frames_without_wrap(source, x);
f = MIN(remaining_frames, fmax);
fmax = audio_stream_frames_without_wrap(sink, y);
f = MIN(f, fmax);
for (j = 0; j < f; j += 2) {
for (i = 0; i < num_filters; i++) {
is = cd->input_channel_select[i];
om = cd->output_channel_mix[i];
s0 = x[is];
s1 = x[is + in_nch];
for (k = 0; k < out_nch; k++) {
if (om & 1) {
y[k] = s0;
y[k + out_nch] = s1;
}
om = om >> 1;
}
}
x += 2 * in_nch;
y += 2 * out_nch;
}
remaining_frames -= f;
x = audio_stream_wrap(source, x);
y = audio_stream_wrap(sink, y);
}
}
#endif
static inline int set_pass_func(struct processing_module *mod, enum sof_ipc_frame fmt)
{
struct tdfb_comp_data *cd = module_get_private_data(mod);
switch (fmt) {
#if CONFIG_FORMAT_S16LE
case SOF_IPC_FRAME_S16_LE:
comp_dbg(mod->dev, "set_pass_func(), SOF_IPC_FRAME_S16_LE");
cd->tdfb_func = tdfb_pass_s16;
break;
#endif /* CONFIG_FORMAT_S16LE */
#if CONFIG_FORMAT_S24LE
case SOF_IPC_FRAME_S24_4LE:
comp_dbg(mod->dev, "set_pass_func(), SOF_IPC_FRAME_S24_4LE");
cd->tdfb_func = tdfb_pass_s24;
break;
#endif /* CONFIG_FORMAT_S24LE */
#if CONFIG_FORMAT_S32LE
case SOF_IPC_FRAME_S32_LE:
comp_dbg(mod->dev, "set_pass_func(), SOF_IPC_FRAME_S32_LE");
cd->tdfb_func = tdfb_pass_s32;
break;
#endif /* CONFIG_FORMAT_S32LE */
default:
comp_err(mod->dev, "set_pass_func(), invalid frame_fmt");
return -EINVAL;
}
return 0;
}
/*
* Control code functions next. The processing is in fir_ C modules.
*/
@ -301,11 +482,25 @@ static void tdfb_init_delay(struct tdfb_comp_data *cd)
}
}
static int tdfb_setup(struct processing_module *mod, int source_nch, int sink_nch)
static int tdfb_setup(struct processing_module *mod, int source_nch, int sink_nch,
enum sof_ipc_frame fmt)
{
struct tdfb_comp_data *cd = module_get_private_data(mod);
int delay_size;
/* If beam on, restore processing function. If off, use for same source and
* sink format the efficient 1:1 copy, otherwise faster pass-through processing
* functions those copy selected source channels to selected sink channels.
*/
if (cd->beam_on) {
set_func(mod, fmt);
} else {
if (source_nch == sink_nch)
cd->tdfb_func = tdfb_pass_same_format;
else
set_pass_func(mod, fmt);
}
/* Set coefficients for each channel from coefficient blob */
delay_size = tdfb_init_coef(mod, source_nch, sink_nch);
if (delay_size < 0)
@ -468,7 +663,8 @@ static int tdfb_process(struct processing_module *mod,
if (comp_is_new_data_blob_available(cd->model_handler)) {
cd->config = comp_get_data_blob(cd->model_handler, NULL, NULL);
ret = tdfb_setup(mod, audio_stream_get_channels(source),
audio_stream_get_channels(sink));
audio_stream_get_channels(sink),
audio_stream_get_frm_fmt(source));
if (ret < 0) {
comp_err(dev, "tdfb_process(), failed FIR setup");
return ret;
@ -479,7 +675,8 @@ static int tdfb_process(struct processing_module *mod,
if (cd->update) {
cd->update = false;
ret = tdfb_setup(mod, audio_stream_get_channels(source),
audio_stream_get_channels(sink));
audio_stream_get_channels(sink),
audio_stream_get_frm_fmt(source));
if (ret < 0) {
comp_err(dev, "tdfb_process(), failed FIR setup");
return ret;
@ -560,7 +757,7 @@ static int tdfb_prepare(struct processing_module *mod,
goto out;
}
ret = tdfb_setup(mod, source_channels, sink_channels);
ret = tdfb_setup(mod, source_channels, sink_channels, frame_fmt);
if (ret < 0) {
comp_err(dev, "tdfb_prepare() error: tdfb_setup failed.");
goto out;