Audio: DRC: DRC math function optimiazation #8431
Conversation
|
here is the xtensa simulator result: and test with xt-testbench 32bit 48kHz on tgl, before optimization we get 150.05mcps(including many trace print and module adapter operation) and after optimization we get 149.13mcps, save about 0.92mcps for component DRC. |
| ae_f32 exp; /* Q7.25 */ | ||
| ae_f32 acc; /* Q6.26 */ | ||
| ae_f32 tmp; /* Q6.26 */ | ||
| ae_f64 tmp64; |
There was a problem hiding this comment.
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.
There was a problem hiding this comment.
I tested in xtensa test bench, it really costs more cycles when using the function wrapper than instructions.
src/audio/drc/drc_math_hifi3.c
Outdated
| x = drc_mult_lshift(x, ONE_OVER_SQRT2_Q30, lshift); | ||
| tmp64 = AE_MULF32R_LL(x, ONE_OVER_SQRT2_Q30); | ||
| /* drc_get_lshift(30, 30, 30) = 1 */ | ||
| tmp64 = AE_SLAI64S(tmp64, 1); |
There was a problem hiding this comment.
You could use #define macro for the magic shift values to know the are for a Qx to Qy conversion.
There was a problem hiding this comment.
if use macro then I can not use AE_SLAI64S, and if use AE_SLAA64S, the log10_fixed reduce cycles from 12.8% down to 12.2%, so better get back the lshift calculation for better code readability.
src/audio/drc/drc_math_hifi3.c
Outdated
| int32_t lshift; | ||
| int32_t e; | ||
| /* drc_get_lshift(25, 30, 25) = 1 */ | ||
| int32_t lshift = 1; |
There was a problem hiding this comment.
Use a #define macro for Qx Qy multiply as Qz shift value?
Use the xtense intrinsic instrunctions directly can save at least 10% cycles for those functions, and save about 0.9mcps for DRC component. Signed-off-by: Andrula Song <andrula.song@intel.com>
|
@andrula-song whats the reason for close if saving MCPS ? |
|
5 participants
Use the xtense intrinsic instrunctions directly can save at least 10% cycles for those functions, and save about 0.92mcps for DRC component.