mirror of
https://github.com/mpv-player/mpv
synced 2025-03-09 15:47:44 +00:00
af_scaletempo: optimize overlap search
scaletempo2 has this optimization where it first uses a step size of 5 together with a quadratic interpolation to quickly get the approximate position of the best overlap and then does a more thorough search aroun that area. Doing the same thing in scaletempo brought a 4.8x performance improvement, however in my measurements a step size of 3 more consistently finds good overlaps and it's still a 2.9x improvement for this function. I should note that while a step size of 3 produced better numbers, I was not actually able to hear any difference in my test. A step size of 3 was chosen just in case it actually makes an audible difference in some cases and the cpu usage isn't really a problem anymore, but that can be revisited in the future. scaletempo2 is still faster then scaletempo with a step size of 5, which I suspect is mostly because it uses some vectorized functions and scaletempo does not.
This commit is contained in:
parent
18ed9e105a
commit
e166ae0ed8
@ -134,50 +134,144 @@ static bool fill_queue(struct priv *s)
|
||||
return bytes_needed == 0;
|
||||
}
|
||||
|
||||
// Fit the curve f(x) = a * x^2 + b * x + c such that
|
||||
// f(-1) = y[0]
|
||||
// f(0) = y[1]
|
||||
// f(1) = y[2]
|
||||
// and return the extremum position and value
|
||||
// assuming y[0] <= y[1] >= y[2] || y[0] >= y[1] <= y[2]
|
||||
static void quadratic_interpolation_float(
|
||||
const float* y_values, float* x, float* value)
|
||||
{
|
||||
const float b = (y_values[2] - y_values[0]) * 0.5f;
|
||||
const float c = y_values[1];
|
||||
const float a = y_values[0] + b - c;
|
||||
|
||||
if (a == 0.f) {
|
||||
// it's a flat line
|
||||
*x = 0;
|
||||
*value = c;
|
||||
} else {
|
||||
const float pos = -b / (2.f * a);
|
||||
*x = pos;
|
||||
*value = a * pos * pos + b * pos + c;
|
||||
}
|
||||
}
|
||||
|
||||
static void quadratic_interpolation_s16(
|
||||
const int32_t* y_values, float* x, int32_t* value)
|
||||
{
|
||||
const float b = (y_values[2] - y_values[0]) * 0.5f;
|
||||
const float c = y_values[1];
|
||||
const float a = y_values[0] + b - c;
|
||||
|
||||
if (a == 0.f) {
|
||||
// it's a flat line
|
||||
*x = 0;
|
||||
*value = c;
|
||||
} else {
|
||||
const float pos = -b / (2.f * a);
|
||||
*x = pos;
|
||||
*value = a * pos * pos + b * pos + c;
|
||||
}
|
||||
}
|
||||
|
||||
static int best_overlap_offset_float(struct priv *s)
|
||||
{
|
||||
float best_distance = FLT_MAX;
|
||||
int best_off = 0;
|
||||
int num_channels = s->num_channels, frames_search = s->frames_search;
|
||||
float *source = (float *)s->buf_queue + num_channels;
|
||||
float *target = (float *)s->buf_overlap + num_channels;
|
||||
int num_samples = s->samples_overlap - num_channels;
|
||||
int step_size = 3;
|
||||
float history[3] = {};
|
||||
|
||||
float *search_start = (float *)s->buf_queue + s->num_channels;
|
||||
for (int off = 0; off < s->frames_search; off++) {
|
||||
float best_distance = FLT_MAX;
|
||||
int best_offset_approx = 0;
|
||||
for (int offset = 0; offset < frames_search; offset += step_size) {
|
||||
float distance = 0;
|
||||
float *ps = search_start;
|
||||
float *po = s->buf_overlap;
|
||||
po += s->num_channels;
|
||||
for (int i = s->num_channels; i < s->samples_overlap; i++)
|
||||
distance += fabsf(*po++ - *ps++);
|
||||
for (int i = 0; i < num_samples; i++)
|
||||
distance += fabsf(target[i] - source[offset * num_channels + i]);
|
||||
|
||||
int offset_approx = offset;
|
||||
history[0] = history[1];
|
||||
history[1] = history[2];
|
||||
history[2] = distance;
|
||||
if(offset >= 2 && history[0] >= history[1] && history[1] <= history[2]) {
|
||||
float extremum;
|
||||
quadratic_interpolation_float(history, &extremum, &distance);
|
||||
offset_approx = offset - step_size + (int)(extremum * step_size + 0.5f);
|
||||
}
|
||||
|
||||
if (distance < best_distance) {
|
||||
best_distance = distance;
|
||||
best_off = off;
|
||||
best_offset_approx = offset_approx;
|
||||
}
|
||||
search_start += s->num_channels;
|
||||
}
|
||||
|
||||
return best_off * 4 * s->num_channels;
|
||||
best_distance = FLT_MAX;
|
||||
int best_offset = 0;
|
||||
int min_offset = MPMAX(0, best_offset_approx - step_size + 1);
|
||||
int max_offset = MPMIN(frames_search, best_offset_approx + step_size);
|
||||
for (int offset = min_offset; offset < max_offset; offset++) {
|
||||
float distance = 0;
|
||||
for (int i = 0; i < num_samples; i++)
|
||||
distance += fabsf(target[i] - source[offset * num_channels + i]);
|
||||
if (distance < best_distance) {
|
||||
best_distance = distance;
|
||||
best_offset = offset;
|
||||
}
|
||||
}
|
||||
|
||||
return best_offset * 4 * num_channels;
|
||||
}
|
||||
|
||||
static int best_overlap_offset_s16(struct priv *s)
|
||||
{
|
||||
int32_t best_distance = INT32_MAX;
|
||||
int best_off = 0;
|
||||
int num_channels = s->num_channels, frames_search = s->frames_search;
|
||||
int16_t *source = (int16_t *)s->buf_queue + num_channels;
|
||||
int16_t *target = (int16_t *)s->buf_overlap + num_channels;
|
||||
int num_samples = s->samples_overlap - num_channels;
|
||||
int step_size = 3;
|
||||
int32_t history[3] = {};
|
||||
|
||||
int16_t *search_start = (int16_t *)s->buf_queue + s->num_channels;
|
||||
for (int off = 0; off < s->frames_search; off++) {
|
||||
int32_t best_distance = INT32_MAX;
|
||||
int best_offset_approx = 0;
|
||||
for (int offset = 0; offset < frames_search; offset += step_size) {
|
||||
int32_t distance = 0;
|
||||
int16_t *ps = search_start;
|
||||
int16_t *po = s->buf_overlap;
|
||||
po += s->num_channels;
|
||||
for (int i = s->num_channels; i < s->samples_overlap; i++)
|
||||
distance += abs((int32_t)*po++ - (int32_t)*ps++);
|
||||
for (int i = 0; i < num_samples; i++)
|
||||
distance += abs((int32_t)target[i] - source[offset * num_channels + i]);
|
||||
|
||||
int offset_approx = offset;
|
||||
history[0] = history[1];
|
||||
history[1] = history[2];
|
||||
history[2] = distance;
|
||||
if(offset >= 2 && history[0] >= history[1] && history[1] <= history[2]) {
|
||||
float extremum;
|
||||
quadratic_interpolation_s16(history, &extremum, &distance);
|
||||
offset_approx = offset - step_size + (int)(extremum * step_size + 0.5f);
|
||||
}
|
||||
|
||||
if (distance < best_distance) {
|
||||
best_distance = distance;
|
||||
best_off = off;
|
||||
best_offset_approx = offset_approx;
|
||||
}
|
||||
search_start += s->num_channels;
|
||||
}
|
||||
|
||||
return best_off * 2 * s->num_channels;
|
||||
best_distance = INT32_MAX;
|
||||
int best_offset = 0;
|
||||
int min_offset = MPMAX(0, best_offset_approx - step_size + 1);
|
||||
int max_offset = MPMIN(frames_search, best_offset_approx + step_size);
|
||||
for (int offset = min_offset; offset < max_offset; offset++) {
|
||||
int32_t distance = 0;
|
||||
for (int i = 0; i < num_samples; i++)
|
||||
distance += abs((int32_t)target[i] - source[offset * num_channels + i]);
|
||||
if (distance < best_distance) {
|
||||
best_distance = distance;
|
||||
best_offset = offset;
|
||||
}
|
||||
}
|
||||
|
||||
return best_offset * 2 * s->num_channels;
|
||||
}
|
||||
|
||||
static void output_overlap_float(struct priv *s, void *buf_out,
|
||||
|
Loading…
Reference in New Issue
Block a user