More ticket #774: optimization for siren7/siren14 codecs
git-svn-id: https://svn.pjsip.org/repos/pjproject/trunk@2616 74dad513-b988-da41-8d7b-12977e46ad98
diff --git a/third_party/g7221/decode/dct4_s.c b/third_party/g7221/decode/dct4_s.c
index 82405a3..0123a13 100644
--- a/third_party/g7221/decode/dct4_s.c
+++ b/third_party/g7221/decode/dct4_s.c
@@ -140,9 +140,9 @@
/*===========================================================*/
/* set_span = 1 << (DCT_LENGTH_LOG - set_count_log); */
- set_span = shr(dct_length,set_count_log);
+ set_span = shr_nocheck(dct_length,set_count_log);
- set_count = shl(1,set_count_log);
+ set_count = shl_nocheck(1,set_count_log);
in_ptr = in_buffer;
move16();
next_out_base = out_buffer;
@@ -185,12 +185,18 @@
/* IF THIS WORKS, IT'S PREFERABLE */
dummy = add(in_val_low,dither_ptr[i++]);
- acca = L_add(dummy,in_val_high);
- out_val_low = extract_l(L_shr(acca,1));
+ // blp: addition of two 16bits vars, there's no way
+ // they'll overflow a 32bit var
+ //acca = L_add(dummy,in_val_high);
+ acca = dummy + in_val_high;
+ out_val_low = extract_l(L_shr_nocheck(acca,1));
dummy = add(in_val_low,dither_ptr[i++]);
- acca = L_add(dummy,-in_val_high);
- out_val_high = extract_l(L_shr(acca,1));
+ // blp: addition of two 16bits vars, there's no way
+ // they'll overflow a 32bit var
+ //acca = L_add(dummy,-in_val_high);
+ acca = dummy - in_val_high;
+ out_val_high = extract_l(L_shr_nocheck(acca,1));
*out_ptr_low++ = out_val_low;
move16();
@@ -284,6 +290,17 @@
{
for ( k=0; k<CORE_SIZE; k++ )
{
+#if PJ_HAS_INT64
+ /* blp: danger danger! not really compatible but faster */
+ pj_int64_t sum64=0;
+ move32();
+
+ for ( i=0; i<CORE_SIZE; i++ )
+ {
+ sum64 += L_mult(pair_ptr[i], dct_core_s[i][k]);
+ }
+ sum = L_saturate(sum64);
+#else
sum=0L;
move32();
@@ -291,6 +308,7 @@
{
sum = L_mac(sum, pair_ptr[i],dct_core_s[i][k]);
}
+#endif
buffer_swap[k] = itu_round(sum);
}
@@ -323,9 +341,9 @@
/*===========================================================*/
/* set_span = 1 << (DCT_LENGTH_LOG - set_count_log); */
- set_span = shr(dct_length,set_count_log);
+ set_span = shr_nocheck(dct_length,set_count_log);
- set_count = shl(1,set_count_log);
+ set_count = shl_nocheck(1,set_count_log);
next_in_base = in_buffer;
move16();
test();
@@ -354,7 +372,7 @@
in_ptr_low = next_in_base;
move16();
- temp = shr(set_span,1);
+ temp = shr_nocheck(set_span,1);
in_ptr_high = in_ptr_low + temp;
move16();
@@ -401,25 +419,25 @@
sum = L_mac(sum,cos_even,in_low_even);
sum = L_mac(sum,negate(msin_even),in_high_even);
- out_low_even = itu_round(L_shl(sum,1));
+ out_low_even = itu_round(L_shl_nocheck(sum,1));
sum = 0L;
move32();
sum = L_mac(sum,msin_even,in_low_even);
sum = L_mac(sum,cos_even,in_high_even);
- out_high_even = itu_round(L_shl(sum,1));
+ out_high_even = itu_round(L_shl_nocheck(sum,1));
sum = 0L;
move32();
sum = L_mac(sum,cos_odd,in_low_odd);
sum = L_mac(sum,msin_odd,in_high_odd);
- out_low_odd = itu_round(L_shl(sum,1));
+ out_low_odd = itu_round(L_shl_nocheck(sum,1));
sum = 0L;
move32();
sum = L_mac(sum,msin_odd,in_low_odd);
sum = L_mac(sum,negate(cos_odd),in_high_odd);
- out_high_odd = itu_round(L_shl(sum,1));
+ out_high_odd = itu_round(L_shl_nocheck(sum,1));
*out_ptr_low++ = out_low_even;
move16();
@@ -458,7 +476,10 @@
{
for(i=0;i<320;i++)
{
- sum = L_add(output[i],syn_bias_7khz[i]);
+ // blp: addition of two 16bits vars, there's no way
+ // they'll overflow a 32bit var
+ //sum = L_add(output[i],syn_bias_7khz[i]);
+ sum = output[i] + syn_bias_7khz[i];
acca = L_sub(sum,32767);
test();
if (acca > 0)
@@ -466,7 +487,10 @@
sum = 32767L;
move32();
}
- acca = L_add(sum,32768L);
+ // blp: addition of two 16bits vars, there's no way
+ // they'll overflow 32bit var
+ //acca = L_add(sum,32768L);
+ acca = sum + 32768;
test();
if (acca < 0)
{