Skip to content

Audio: DRC: DRC math function optimiazation #8431

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
andrula-song wants to merge 1 commit into from
Closed

Audio: DRC: DRC math function optimiazation #8431

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
138 changes: 92 additions & 46 deletions src/audio/drc/drc_math_hifi3.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,19 +38,7 @@
#define INV_FUNC_A2_Q25 1126492160 /* Q_CONVERT_FLOAT(33.57208251953125f, 25) */
#define INV_FUNC_A1_Q25 -713042175 /* Q_CONVERT_FLOAT(-21.25031280517578125f, 25) */
#define INV_FUNC_A0_Q25 239989712 /* Q_CONVERT_FLOAT(7.152250766754150390625f, 25) */

/*
* Input depends on precision_x
* Output range [0.5, 1); regulated to Q2.30
*/
static inline ae_f32 rexp_fixed(ae_f32 x, int32_t precision_x, int32_t *e)
{
int32_t bit = 31 - AE_NSAZ32_L(x);

*e = bit - precision_x;

return AE_SRAA32(x, bit - 30);
}
#define DRC_TWENTY_Q26 1342177280 /* Q_CONVERT_FLOAT(20, 26) */

/*
* Input is Q6.26: max 32.0
Expand All @@ -62,35 +50,52 @@ static inline ae_f32 log10_fixed(ae_f32 x)
* fpminimax(log10(x), 5, [|SG...|], [1/2;sqrt(2)/2], absolute);
* max err ~= 6.088e-8
*/
int32_t e;
int32_t lshift;
ae_f32 exp; /* Q7.25 */
ae_f32 acc; /* Q6.26 */
ae_f32 tmp; /* Q6.26 */
ae_f64 tmp64;
Copy link
Collaborator

@singalsu singalsu Nov 3, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So, there's overhead in inline functions, maybe it's from 26 as literal instead of variable? You could comment that the instructions normalize the value after the function was removed.

Copy link
Contributor Author

@andrula-song andrula-song Nov 6, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I tested in xtensa test bench, it really costs more cycles when using the function wrapper than instructions.

int bit = 31 - AE_NSAZ32_L(x);
int32_t e = bit - 26;
int32_t lshift;

x = rexp_fixed(x, 26, &e); /* Q2.30 */
x = AE_SRAA32(x, bit - 30); /* Q2.30 */
exp = e << 25; /* Q_CONVERT_FLOAT(e, 25) */

if ((int32_t)x > (int32_t)ONE_OVER_SQRT2_Q30) {
tmp64 = AE_MULF32R_LL(x, ONE_OVER_SQRT2_Q30);
/* drc_get_lshift(30, 30, 30) = 1 */
lshift = drc_get_lshift(30, 30, 30);
x = drc_mult_lshift(x, ONE_OVER_SQRT2_Q30, lshift);
tmp64 = AE_SLAA64S(tmp64, lshift);
x = AE_ROUND32F48SSYM(tmp64);
exp = AE_ADD32(exp, HALF_Q25);
}

tmp64 = AE_MULF32R_LL(LOG10_FUNC_A5_Q26, x);
/* drc_get_lshift(26, 30, 26) = 1 */
lshift = drc_get_lshift(26, 30, 26);
acc = drc_mult_lshift(LOG10_FUNC_A5_Q26, x, lshift);
tmp64 = AE_SLAA64S(tmp64, lshift);
acc = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, LOG10_FUNC_A4_Q26);
acc = drc_mult_lshift(acc, x, lshift);
tmp64 = AE_MULF32R_LL(acc, x);
tmp64 = AE_SLAA64S(tmp64, lshift);
acc = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, LOG10_FUNC_A3_Q26);
acc = drc_mult_lshift(acc, x, lshift);
tmp64 = AE_MULF32R_LL(acc, x);
tmp64 = AE_SLAA64S(tmp64, lshift);
acc = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, LOG10_FUNC_A2_Q26);
acc = drc_mult_lshift(acc, x, lshift);
tmp64 = AE_MULF32R_LL(acc, x);
tmp64 = AE_SLAA64S(tmp64, lshift);
acc = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, LOG10_FUNC_A1_Q26);
acc = drc_mult_lshift(acc, x, lshift);
tmp64 = AE_MULF32R_LL(acc, x);
tmp64 = AE_SLAA64S(tmp64, lshift);
acc = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, LOG10_FUNC_A0_Q26);

tmp64 = AE_MULF32R_LL(exp, LOG10_2_Q26);
lshift = drc_get_lshift(25, 26, 26);
tmp = drc_mult_lshift(exp, LOG10_2_Q26, lshift);
tmp64 = AE_SLAA64S(tmp64, lshift);
tmp = AE_ROUND32F48SSYM(tmp64);
acc = AE_ADD32(acc, tmp);

return acc;
Expand All @@ -103,13 +108,19 @@ static inline ae_f32 log10_fixed(ae_f32 x)
inline int32_t drc_lin2db_fixed(int32_t linear)
{
ae_f32 log10_linear;
ae_f32 y;
ae_f64 tmp;

/* For negative or zero, just return a very small dB value. */
if (linear <= 0)
return NEG_1K_Q21;

log10_linear = log10_fixed(linear); /* Q6.26 */
return drc_mult_lshift(20 << 26, log10_linear, drc_get_lshift(26, 26, 21));

tmp = AE_MULF32R_LL(DRC_TWENTY_Q26, log10_linear);
/*Don't need to do AE_SLAA64S since drc_get_lshift(26, 26, 21) = 0 */
y = AE_ROUND32F48SSYM(tmp);
return y;
}

/*
Expand All @@ -119,13 +130,18 @@ inline int32_t drc_lin2db_fixed(int32_t linear)
inline int32_t drc_log_fixed(int32_t x)
{
ae_f32 log10_x;
ae_f64 tmp;
ae_f32 y;

if (x <= 0)
return NEG_30_Q26;

/* log(x) = log(10) * log10(x) */
log10_x = log10_fixed(x); /* Q6.26 */
return drc_mult_lshift(LOG_10_Q29, log10_x, drc_get_lshift(29, 26, 26));
tmp = AE_MULF32R_LL(LOG_10_Q29, log10_x);
tmp = AE_SLAA64S(tmp, drc_get_lshift(29, 26, 26));
y = AE_ROUND32F48SSYM(tmp);
return y;
}

#ifndef DRC_USE_CORDIC_ASIN
Expand All @@ -143,17 +159,18 @@ inline int32_t drc_asin_fixed(int32_t x)
* fpminimax(asin(x), [|1,3,5,7|], [|SG...|], [1/sqrt(2);1], absolute)
* max err ~= 3.085226e-2
*/
int32_t lshift;
ae_f32 in = x; /* Q2.30 */
int32_t lshift = drc_get_lshift(30, 30, 30);
ae_f32 in2; /* Q2.30 */
ae_f32 A7, A5, A3, A1;
int32_t qc;
ae_f32 acc;
ae_f64 tmp;

lshift = drc_get_lshift(30, 30, 30);
in2 = drc_mult_lshift(in, in, lshift);
tmp = AE_MULF32R_LL(x, x);
tmp = AE_SLAA64S(tmp, lshift);
in2 = AE_ROUND32F48SSYM(tmp);

if (ABS((int32_t)in) <= ONE_OVER_SQRT2_Q30) {
if (ABS((int32_t)x) <= ONE_OVER_SQRT2_Q30) {
A7 = ASIN_FUNC_A7L_Q30;
A5 = ASIN_FUNC_A5L_Q30;
A3 = ASIN_FUNC_A3L_Q30;
Expand All @@ -168,15 +185,26 @@ inline int32_t drc_asin_fixed(int32_t x)
}

lshift = drc_get_lshift(qc, 30, qc);
acc = drc_mult_lshift(A7, in2, lshift);
tmp = AE_MULF32R_LL(A7, in2);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, A5);
acc = drc_mult_lshift(acc, in2, lshift);
tmp = AE_MULF32R_LL(acc, in2);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, A3);
acc = drc_mult_lshift(acc, in2, lshift);
tmp = AE_MULF32R_LL(acc, in2);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, A1);
acc = drc_mult_lshift(acc, in, lshift);
tmp = AE_MULF32R_LL(acc, x);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
lshift = drc_get_lshift(qc, 30, 30);
return drc_mult_lshift(acc, TWO_OVER_PI_Q30, lshift);
tmp = AE_MULF32R_LL(acc, TWO_OVER_PI_Q30);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
return acc;
}
#endif /* !DRC_USE_CORDIC_ASIN */

Expand All @@ -191,34 +219,52 @@ inline int32_t drc_inv_fixed(int32_t x, int32_t precision_x, int32_t precision_y
* max err ~= 1.00388e-6
*/
ae_f32 in;
/* drc_get_lshift(25, 30, 25) = 1 */
int32_t lshift;
int32_t e;
int32_t precision_inv;
int32_t sqrt2_extracted = 0;
ae_f32 acc;
ae_f64 tmp;
int32_t bit = 31 - AE_NSAZ32_L(x);
int32_t e = bit - precision_x;

in = rexp_fixed(x, precision_x, &e); /* Q2.30 */
in = AE_SRAA32(x, bit - 30); /* Q2.30 */

if (ABS((int32_t)in) < ONE_OVER_SQRT2_Q30) {
lshift = drc_get_lshift(30, 30, 30);
in = drc_mult_lshift(in, SQRT2_Q30, lshift);
tmp = AE_MULF32R_LL(in, SQRT2_Q30);
tmp = AE_SLAA64S(tmp, lshift);
in = AE_ROUND32F48SSYM(tmp);
sqrt2_extracted = 1;
}

lshift = drc_get_lshift(25, 30, 25);
acc = drc_mult_lshift(INV_FUNC_A5_Q25, in, lshift);
tmp = AE_MULF32R_LL(in, INV_FUNC_A5_Q25);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, INV_FUNC_A4_Q25);
acc = drc_mult_lshift(acc, in, lshift);
tmp = AE_MULF32R_LL(in, acc);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, INV_FUNC_A3_Q25);
acc = drc_mult_lshift(acc, in, lshift);
tmp = AE_MULF32R_LL(in, acc);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, INV_FUNC_A2_Q25);
acc = drc_mult_lshift(acc, in, lshift);
tmp = AE_MULF32R_LL(in, acc);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, INV_FUNC_A1_Q25);
acc = drc_mult_lshift(acc, in, lshift);
tmp = AE_MULF32R_LL(in, acc);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
acc = AE_ADD32(acc, INV_FUNC_A0_Q25);

if (sqrt2_extracted)
acc = drc_mult_lshift(acc, SQRT2_Q30, lshift);
if (sqrt2_extracted) {
tmp = AE_MULF32R_LL(SQRT2_Q30, acc);
tmp = AE_SLAA64S(tmp, lshift);
acc = AE_ROUND32F48SSYM(tmp);
}

precision_inv = e + 25;
acc = AE_SLAA32S(acc, precision_y - precision_inv);
Expand Down