Optimize ssim_8x8 for SSE2.
Code was pulled from libvpx assembly and converted to intrinsics. Review URL: http://webrtc-codereview.appspot.com/122001 git-svn-id: http://webrtc.googlecode.com/svn/trunk@409 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
frkoenig@google.com
parent
6423509efc
commit
9139fddf0e
@ -8,6 +8,7 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
#include "system_wrappers/interface/cpu_features_wrapper.h"
|
||||
#include "test_util.h"
|
||||
#include "test_macros.h"
|
||||
#include "rtp_dump.h"
|
||||
@ -440,34 +441,6 @@ PSNRfromFiles(const WebRtc_Word8 *refFileName, const WebRtc_Word8 *testFileName,
|
||||
return 0;
|
||||
}
|
||||
|
||||
void
|
||||
ssim_parms_8x8_c
|
||||
(
|
||||
unsigned char *s,
|
||||
int sp,
|
||||
unsigned char *r,
|
||||
int rp,
|
||||
unsigned long *sum_s,
|
||||
unsigned long *sum_r,
|
||||
unsigned long *sum_sq_s,
|
||||
unsigned long *sum_sq_r,
|
||||
unsigned long *sum_sxr
|
||||
)
|
||||
{
|
||||
int i,j;
|
||||
for(i=0;i<8;i++,s+=sp,r+=rp)
|
||||
{
|
||||
for(j=0;j<8;j++)
|
||||
{
|
||||
*sum_s += s[j];
|
||||
*sum_r += r[j];
|
||||
*sum_sq_s += s[j] * s[j];
|
||||
*sum_sq_r += r[j] * r[j];
|
||||
*sum_sxr += s[j] * r[j];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static double
|
||||
similarity
|
||||
(
|
||||
@ -499,13 +472,110 @@ similarity
|
||||
}
|
||||
|
||||
static double
|
||||
ssim_8x8(unsigned char *s,int sp, unsigned char *r,int rp)
|
||||
ssim_8x8_c
|
||||
(
|
||||
unsigned char *s,
|
||||
int sp,
|
||||
unsigned char *r,
|
||||
int rp
|
||||
)
|
||||
{
|
||||
unsigned long sum_s=0,sum_r=0,sum_sq_s=0,sum_sq_r=0,sum_sxr=0;
|
||||
ssim_parms_8x8_c(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r, &sum_sxr);
|
||||
unsigned long sum_s = 0;
|
||||
unsigned long sum_r = 0;
|
||||
unsigned long sum_sq_s = 0;
|
||||
unsigned long sum_sq_r = 0;
|
||||
unsigned long sum_sxr = 0;
|
||||
|
||||
int i,j;
|
||||
for(i=0;i<8;i++,s+=sp,r+=rp)
|
||||
{
|
||||
for(j=0;j<8;j++)
|
||||
{
|
||||
sum_s += s[j];
|
||||
sum_r += r[j];
|
||||
sum_sq_s += s[j] * s[j];
|
||||
sum_sq_r += r[j] * r[j];
|
||||
sum_sxr += s[j] * r[j];
|
||||
}
|
||||
}
|
||||
|
||||
return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64);
|
||||
}
|
||||
|
||||
#if defined(WEBRTC_USE_SSE2)
|
||||
#include <xmmintrin.h>
|
||||
static double
|
||||
ssim_8x8_sse2
|
||||
(
|
||||
unsigned char *s,
|
||||
int sp,
|
||||
unsigned char *r,
|
||||
int rp
|
||||
)
|
||||
{
|
||||
int i;
|
||||
const __m128i z = _mm_setzero_si128();
|
||||
__m128i sum_s_16 = _mm_setzero_si128();
|
||||
__m128i sum_r_16 = _mm_setzero_si128();
|
||||
__m128i sum_sq_s_32 = _mm_setzero_si128();
|
||||
__m128i sum_sq_r_32 = _mm_setzero_si128();
|
||||
__m128i sum_sxr_32 = _mm_setzero_si128();
|
||||
|
||||
for(i=0;i<8;i++,s+=sp,r+=rp)
|
||||
{
|
||||
const __m128i s_8 = _mm_loadl_epi64((__m128i*)(s));
|
||||
const __m128i r_8 = _mm_loadl_epi64((__m128i*)(r));
|
||||
|
||||
const __m128i s_16 = _mm_unpacklo_epi8(s_8,z);
|
||||
const __m128i r_16 = _mm_unpacklo_epi8(r_8,z);
|
||||
|
||||
sum_s_16 = _mm_adds_epu16(sum_s_16, s_16);
|
||||
sum_r_16 = _mm_adds_epu16(sum_r_16, r_16);
|
||||
const __m128i sq_s_32 = _mm_madd_epi16(s_16, s_16);
|
||||
sum_sq_s_32 = _mm_add_epi32(sum_sq_s_32, sq_s_32);
|
||||
const __m128i sq_r_32 = _mm_madd_epi16(r_16, r_16);
|
||||
sum_sq_r_32 = _mm_add_epi32(sum_sq_r_32, sq_r_32);
|
||||
const __m128i sxr_32 = _mm_madd_epi16(s_16, r_16);
|
||||
sum_sxr_32 = _mm_add_epi32(sum_sxr_32, sxr_32);
|
||||
}
|
||||
|
||||
const __m128i sum_s_32 = _mm_add_epi32(_mm_unpackhi_epi16(sum_s_16, z),
|
||||
_mm_unpacklo_epi16(sum_s_16, z));
|
||||
const __m128i sum_r_32 = _mm_add_epi32(_mm_unpackhi_epi16(sum_r_16, z),
|
||||
_mm_unpacklo_epi16(sum_r_16, z));
|
||||
|
||||
unsigned long sum_s_64[2];
|
||||
unsigned long sum_r_64[2];
|
||||
unsigned long sum_sq_s_64[2];
|
||||
unsigned long sum_sq_r_64[2];
|
||||
unsigned long sum_sxr_64[2];
|
||||
|
||||
_mm_store_si128 ((__m128i*)sum_s_64,
|
||||
_mm_add_epi64(_mm_unpackhi_epi32(sum_s_32, z),
|
||||
_mm_unpacklo_epi32(sum_s_32, z)));
|
||||
_mm_store_si128 ((__m128i*)sum_r_64,
|
||||
_mm_add_epi64(_mm_unpackhi_epi32(sum_r_32, z),
|
||||
_mm_unpacklo_epi32(sum_r_32, z)));
|
||||
_mm_store_si128 ((__m128i*)sum_sq_s_64,
|
||||
_mm_add_epi64(_mm_unpackhi_epi32(sum_sq_s_32, z),
|
||||
_mm_unpacklo_epi32(sum_sq_s_32, z)));
|
||||
_mm_store_si128 ((__m128i*)sum_sq_r_64,
|
||||
_mm_add_epi64(_mm_unpackhi_epi32(sum_sq_r_32, z),
|
||||
_mm_unpacklo_epi32(sum_sq_r_32, z)));
|
||||
_mm_store_si128 ((__m128i*)sum_sxr_64,
|
||||
_mm_add_epi64(_mm_unpackhi_epi32(sum_sxr_32, z),
|
||||
_mm_unpacklo_epi32(sum_sxr_32, z)));
|
||||
|
||||
const unsigned long sum_s = sum_s_64[0] + sum_s_64[1];
|
||||
const unsigned long sum_r = sum_r_64[0] + sum_r_64[1];
|
||||
const unsigned long sum_sq_s = sum_sq_s_64[0] + sum_sq_s_64[1];
|
||||
const unsigned long sum_sq_r = sum_sq_r_64[0] + sum_sq_r_64[1];
|
||||
const unsigned long sum_sxr = sum_sxr_64[0] + sum_sxr_64[1];
|
||||
|
||||
return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64);
|
||||
}
|
||||
#endif
|
||||
|
||||
double
|
||||
SSIM_frame
|
||||
(
|
||||
@ -516,9 +586,18 @@ SSIM_frame
|
||||
int width,
|
||||
int height)
|
||||
{
|
||||
WebRtc_Word32 i,j;
|
||||
WebRtc_UWord32 samples = 0;
|
||||
int i,j;
|
||||
unsigned int samples = 0;
|
||||
double ssim_total = 0;
|
||||
double (*ssim_8x8)(unsigned char*, int, unsigned char*, int rp);
|
||||
|
||||
ssim_8x8 = ssim_8x8_c;
|
||||
if(WebRtc_GetCPUInfo(kSSE2))
|
||||
{
|
||||
#if defined(WEBRTC_USE_SSE2)
|
||||
ssim_8x8 = ssim_8x8_sse2;
|
||||
#endif
|
||||
}
|
||||
|
||||
// sample point start with each 4x4 location
|
||||
for(i=0; i < height-8; i+=4, img1 += stride_img1*4, img2 += stride_img2*4)
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user