[x265] [PATCH] asm: fix sse_ss for [16x16], [32x32] and [64x64] sse2 12bpp

[chen]

At 2015-09-18 12:30:06,ramya at multicorewareinc.com wrote:
># HG changeset patch
># User Ramya Sriraman <ramya at multicorewareinc.com>
># Date 1442550588 -19800
>#      Fri Sep 18 09:59:48 2015 +0530
># Node ID 2cca9810882147d5aece67e22403d5d40f768024
># Parent  8db83511da0b11b7347adea081269e3591029841
>asm: fix sse_ss for [16x16], [32x32] and [64x64] sse2 12bpp
>
>diff -r 8db83511da0b -r 2cca98108821 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Mon Sep 14 09:28:07 2015 +0530
>+++ b/source/common/x86/asm-primitives.cpp	Fri Sep 18 09:59:48 2015 +0530
>@@ -1002,10 +1002,12 @@
>         p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].sse_pp = (pixel_sse_t)PFX(pixel_ssd_ss_8x16_sse2);
>         p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sse_pp = (pixel_sse_t)PFX(pixel_ssd_ss_16x32_sse2);
>         p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sse_pp = (pixel_sse_t)PFX(pixel_ssd_ss_32x64_sse2);
>-#if X265_DEPTH <= 10
>-        p.cu[BLOCK_4x4].sse_ss = PFX(pixel_ssd_ss_4x4_mmx2);
>-        ALL_LUMA_CU(sse_ss, pixel_ssd_ss, sse2);
>-#endif
>+        p.cu[BLOCK_4x4].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_4x4_mmx2);
>+        p.cu[BLOCK_8x8].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_8x8_sse2);
>+        p.cu[BLOCK_16x16].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_16x16_sse2);
>+        p.cu[BLOCK_32x32].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_32x32_sse2);
>+        p.cu[BLOCK_64x64].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_64x64_sse2);
>+
>         p.cu[BLOCK_4x4].dct = PFX(dct4_sse2);
>         p.cu[BLOCK_8x8].dct = PFX(dct8_sse2);
>         p.cu[BLOCK_4x4].idct = PFX(idct4_sse2);
>diff -r 8db83511da0b -r 2cca98108821 source/common/x86/ssd-a.asm
>--- a/source/common/x86/ssd-a.asm	Mon Sep 14 09:28:07 2015 +0530
>+++ b/source/common/x86/ssd-a.asm	Fri Sep 18 09:59:48 2015 +0530
>@@ -181,6 +181,197 @@
>     RET
> %endmacro
> 
>+;Function to find ssd for 8x16 block, sse2, 12 bit depth
>+;Defined sepeartely to be called from SSD_ONE_SS_16 macro
>+INIT_XMM sse2
>+cglobal ssd_ss_8x16
>+    pxor        m8, m8
>+    mov         r4d, 4
>+.loop:
>+    movu        m0, [r0]
>+    movu        m1, [r0 + mmsize]
>+    movu        m2, [r0 + r1]
>+    movu        m3, [r0 + r1+ mmsize]
>+    movu        m4, [r2]
>+    movu        m5, [r2 + mmsize]
>+    movu        m6, [r2 + r3]
>+    movu        m7, [r2 + r3 + mmsize]
>+    psubw       m0, m4
>+    psubw       m1, m5
>+    psubw       m2, m6
>+    psubw       m3, m7
>+    lea         r0, [r0 + 2 * r1]
>+    lea         r2, [r2 + 2 * r3]
>+    pmaddwd     m0, m0
>+    pmaddwd     m1, m1
>+    pmaddwd     m2, m2
>+    pmaddwd     m3, m3
>+    paddd       m2, m3
>+    paddd       m0, m1
>+    paddd       m0, m2
paddd m8, m2  <-- this style broken dependency link, may faster
>+    paddd       m8, m0
>+    dec         r4d
>+    jnz         .loop
>+
>+    mova        m4, m8
>+    pxor        m5, m5
>+    punpckldq   m8, m5
>+    punpckhdq   m4, m5
>+    paddq       m4, m8
>+    movhlps     m5, m4
>+    paddq       m4, m5
>+    paddq       m9, m4
in this case, dynamic range is
12+12 + 3+4 = 31, so we didn't need QWORD operators (3+4 because 8x16)
another problem, this function as common function to share with other functions, so we didn't need sum all elements into one QWORD, we can do it in last stage.
same comment in below code

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150922/ad5457bf/attachment.html>