generic-library/webrtc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

performance in ARM V7 improved 7~8 %.

Browse Source 
Review URL: http://webrtc-codereview.appspot.com/114007

git-svn-id: http://webrtc.googlecode.com/svn/trunk@384 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
kma@google.com 2011-08-16 20:00:18 +00:00
parent 977c2966fc
commit 78dc99e2a1
2 changed files with 49 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
src/common_audio/signal_processing_library/main/source/complex_fft.c
View File

@ -98,19 +98,39 @@ int WebRtcSpl_ComplexFFT(WebRtc_Word16 frfi[], int stages, int mode)
wr = WebRtcSpl_kSinTable1024[j + 256]; wr = WebRtcSpl_kSinTable1024[j + 256];
wi = -WebRtcSpl_kSinTable1024[j]; wi = -WebRtcSpl_kSinTable1024[j];
#ifdef WEBRTC_ARCH_ARM_V7A
WebRtc_Word32 wri;
WebRtc_Word32 frfi_r;
__asm__("pkhbt %0, %1, %2, lsl #16" : "=r"(wri) :
"r"((WebRtc_Word32)wr), "r"((WebRtc_Word32)wi));
#endif
for (i = m; i < n; i += istep) for (i = m; i < n; i += istep)
{ {
j = i + l; j = i + l;
tr32 = WEBRTC_SPL_RSHIFT_W32((WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j]) #ifdef WEBRTC_ARCH_ARM_V7A
- WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j + 1]) + CFFTRND), __asm__("pkhbt %0, %1, %2, lsl #16" : "=r"(frfi_r) :
15 - CFFTSFT); "r"((WebRtc_Word32)frfi[2*j]), "r"((WebRtc_Word32)frfi[2*j +1]));
__asm__("smlsd %0, %1, %2, %3" : "=r"(tr32) :
"r"(wri), "r"(frfi_r), "r"(CFFTRND));
__asm__("smladx %0, %1, %2, %3" : "=r"(ti32) :
"r"(wri), "r"(frfi_r), "r"(CFFTRND));
ti32 = WEBRTC_SPL_RSHIFT_W32((WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j + 1]) #else
+ WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j]) + CFFTRND), 15 - CFFTSFT); tr32 = WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j])
- WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j + 1]) + CFFTRND;
ti32 = WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j + 1])
+ WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j]) + CFFTRND;
#endif
tr32 = WEBRTC_SPL_RSHIFT_W32(tr32, 15 - CFFTSFT);
ti32 = WEBRTC_SPL_RSHIFT_W32(ti32, 15 - CFFTSFT);
qr32 = ((WebRtc_Word32)frfi[2 * i]) << CFFTSFT; qr32 = ((WebRtc_Word32)frfi[2 * i]) << CFFTSFT;
qi32 = ((WebRtc_Word32)frfi[2 * i + 1]) << CFFTSFT; qi32 = ((WebRtc_Word32)frfi[2 * i + 1]) << CFFTSFT;
frfi[2 * j] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32( frfi[2 * j] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(
(qr32 - tr32 + CFFTRND2), 1 + CFFTSFT); (qr32 - tr32 + CFFTRND2), 1 + CFFTSFT);
frfi[2 * j + 1] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32( frfi[2 * j + 1] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(

31
src/common_audio/signal_processing_library/main/source/complex_ifft.c
View File

@ -110,21 +110,38 @@ int WebRtcSpl_ComplexIFFT(WebRtc_Word16 frfi[], int stages, int mode)
wr = WebRtcSpl_kSinTable1024[j + 256]; wr = WebRtcSpl_kSinTable1024[j + 256];
wi = WebRtcSpl_kSinTable1024[j]; wi = WebRtcSpl_kSinTable1024[j];
#ifdef WEBRTC_ARCH_ARM_V7A
WebRtc_Word32 wri;
WebRtc_Word32 frfi_r;
__asm__("pkhbt %0, %1, %2, lsl #16" : "=r"(wri) :
"r"((WebRtc_Word32)wr), "r"((WebRtc_Word32)wi));
#endif
for (i = m; i < n; i += istep) for (i = m; i < n; i += istep)
{ {
j = i + l; j = i + l;
tr32 = WEBRTC_SPL_RSHIFT_W32((WEBRTC_SPL_MUL_16_16_RSFT(wr, frfi[2 * j], 0) #ifdef WEBRTC_ARCH_ARM_V7A
- WEBRTC_SPL_MUL_16_16_RSFT(wi, frfi[2 * j + 1], 0) + CIFFTRND), __asm__("pkhbt %0, %1, %2, lsl #16" : "=r"(frfi_r) :
15 - CIFFTSFT); "r"((WebRtc_Word32)frfi[2*j]), "r"((WebRtc_Word32)frfi[2*j +1]));
__asm__("smlsd %0, %1, %2, %3" : "=r"(tr32) :
"r"(wri), "r"(frfi_r), "r"(CIFFTRND));
__asm__("smladx %0, %1, %2, %3" : "=r"(ti32) :
"r"(wri), "r"(frfi_r), "r"(CIFFTRND));
#else
ti32 = WEBRTC_SPL_RSHIFT_W32( tr32 = WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j])
(WEBRTC_SPL_MUL_16_16_RSFT(wr, frfi[2 * j + 1], 0) - WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j + 1]) + CIFFTRND;
+ WEBRTC_SPL_MUL_16_16_RSFT(wi, frfi[2 * j], 0)
+ CIFFTRND), 15 - CIFFTSFT); ti32 = WEBRTC_SPL_MUL_16_16(wr, frfi[2 * j + 1])
+ WEBRTC_SPL_MUL_16_16(wi, frfi[2 * j]) + CIFFTRND;
#endif
tr32 = WEBRTC_SPL_RSHIFT_W32(tr32, 15 - CIFFTSFT);
ti32 = WEBRTC_SPL_RSHIFT_W32(ti32, 15 - CIFFTSFT);
qr32 = ((WebRtc_Word32)frfi[2 * i]) << CIFFTSFT; qr32 = ((WebRtc_Word32)frfi[2 * i]) << CIFFTSFT;
qi32 = ((WebRtc_Word32)frfi[2 * i + 1]) << CIFFTSFT; qi32 = ((WebRtc_Word32)frfi[2 * i + 1]) << CIFFTSFT;
frfi[2 * j] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((qr32 - tr32+round2), frfi[2 * j] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((qr32 - tr32+round2),
shift+CIFFTSFT); shift+CIFFTSFT);
frfi[2 * j + 1] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32( frfi[2 * j + 1] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(