[x265] [PATCH Review only] asm: added psy_acEnergy_pp_4x4 in sse4 for psyCost_pp

[chen]

At 2014-12-16 18:35:49,"Divya Manivannan" <divya at multicorewareinc.com> wrote:
># HG changeset patch
># User Divya Manivannan <divya at multicorewareinc.com>
># Date 1418726099 -19800
>#      Tue Dec 16 16:04:59 2014 +0530
># Node ID de6f39b44c144aa56c68d27d6ee201e7dd493755
># Parent  775ebb4694ad7931a98b796640bf646085659ea2
>asm: added psy_acEnergy_pp_4x4 in sse4 for psyCost_pp
>
>diff -r 775ebb4694ad -r de6f39b44c14 source/common/pixel.cpp
>--- a/source/common/pixel.cpp	Tue Dec 16 09:40:00 2014 +0530
>+++ b/source/common/pixel.cpp	Tue Dec 16 16:04:59 2014 +0530
>@@ -795,8 +795,18 @@
>     else
>     {
>         /* 4x4 is too small for sa8d */
>-        int sourceEnergy = satd_4x4(source, sstride, zeroBuf, 0) - (sad<4, 4>(source, sstride, zeroBuf, 0) >> 2);
>-        int reconEnergy = satd_4x4(recon, rstride, zeroBuf, 0) - (sad<4, 4>(recon, rstride, zeroBuf, 0) >> 2);
>+        int sourceEnergy, reconEnergy;
>+        if (!HIGH_BIT_DEPTH)    // once HBD asm code is developed, if condition will go away

#if
 
>+        {
>+            sourceEnergy = primitives.psy_acEnergy_pp(source, sstride);
>+            reconEnergy = primitives.psy_acEnergy_pp(recon, rstride);
>+        }
>+        else
>+        {
>+            //original code;
>+            sourceEnergy = satd_4x4(source, sstride, zeroBuf, 0) - (sad<4, 4>(source, sstride, zeroBuf, 0) >> 2);
>+            reconEnergy = satd_4x4(recon, rstride, zeroBuf, 0) - (sad<4, 4>(recon, rstride, zeroBuf, 0) >> 2);

reference code may put into primitives C model
>+        }
>         return abs(sourceEnergy - reconEnergy);
>     }
> }
>diff -r 775ebb4694ad -r de6f39b44c14 source/common/primitives.h
>--- a/source/common/primitives.h	Tue Dec 16 09:40:00 2014 +0530
>+++ b/source/common/primitives.h	Tue Dec 16 16:04:59 2014 +0530
>@@ -195,6 +195,7 @@
> typedef void (*planecopy_sp_t) (const uint16_t* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int width, int height, int shift, uint16_t mask);
> 
> typedef void (*cutree_propagate_cost) (int* dst, const uint16_t* propagateIn, const int32_t* intraCosts, const uint16_t* interCosts, const int32_t* invQscales, const double* fpsFactor, int len);
>+typedef int(*psy_acEnergy_pp_t)(const pixel* pix, intptr_t stride);
> 
> /* Define a structure containing function pointers to optimized encoder
>  * primitives.  Each pointer can reference either an assembly routine,
>@@ -213,6 +214,7 @@
>     pixelcmp_t            sa8d[NUM_SQUARE_BLOCKS];         // sa8d primitives for square intra blocks
>     pixelcmp_t            psy_cost_pp[NUM_SQUARE_BLOCKS];  // difference in AC energy between two blocks
>     pixelcmp_ss_t         psy_cost_ss[NUM_SQUARE_BLOCKS];
>+    psy_acEnergy_pp_t     psy_acEnergy_pp;
> 
>     dct_t                 dct[NUM_DCTS];
>     idct_t                idct[NUM_IDCTS];
>diff -r 775ebb4694ad -r de6f39b44c14 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Tue Dec 16 09:40:00 2014 +0530
>+++ b/source/common/x86/asm-primitives.cpp	Tue Dec 16 16:04:59 2014 +0530
>@@ -1898,6 +1898,9 @@
>         p.chroma[X265_CSP_I420].filter_vpp[CHROMA_16x16] = x265_interp_4tap_vert_pp_16x16_avx2;
>         p.chroma[X265_CSP_I420].filter_vpp[CHROMA_32x32] = x265_interp_4tap_vert_pp_32x32_avx2;
> #endif
>+
>+        p.psy_acEnergy_pp = x265_psy_acEnergy_pp_4x4_sse4;
>+
>     }
> #endif // if HIGH_BIT_DEPTH
> }
>diff -r 775ebb4694ad -r de6f39b44c14 source/common/x86/pixel-a.asm
>--- a/source/common/x86/pixel-a.asm	Tue Dec 16 09:40:00 2014 +0530
>+++ b/source/common/x86/pixel-a.asm	Tue Dec 16 16:04:59 2014 +0530
>@@ -6579,3 +6579,35 @@
>     mov         [r2], r3w
> .end:
>     RET
>+
>+;---------------------------------------------------------------------------------------------------------------------
>+;int psy_acEnergy_pp(const pixel* source, intptr_t sstride)
>+;---------------------------------------------------------------------------------------------------------------------
>+INIT_XMM sse4
>+cglobal psy_acEnergy_pp_4x4, 2, 3, 6
>+
>+    lea             r2, [3 * r1]
>+    movd            m0, [r0]
>+    movd            m1, [r0 + r1]
>+    movd            m2, [r0 + r1 * 2]
>+    movd            m3, [r0 + r2]
>+    shufps          m0, m1, 0

overwrite m0 with m1 lowest 32-bits?
the compute logic is wrong below
 
>+    shufps          m2, m3, 0
>+    mova            m4, [hmul_4p]
>+    pmaddubsw       m0, m4
>+    pmaddubsw       m2, m4
>+
>+    paddw           m5, m0, m2
>+    movhlps         m4, m5
>+    paddw           m5, m4
>+    phaddw          m5, m5

pmaddwd may replace this phaddw to avoid Port5, just another way, need analyze output object code.
 
>+    pmovzxwd        m5, m5

reduce by above pmaddwd
 
>+    psrld           m5, 2
>+
>+    HADAMARD 0, sumsub, 0, 2, 1, 3
>+    HADAMARD 4, sumsub, 0, 2, 1, 3
>+    HADAMARD 1, amax, 0, 2, 1, 3
>+    HADDW m0, m2
>+    psubd           m0, m5
>+    movd            eax, m0
>+    RET
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20141217/e45e9f79/attachment-0001.html>