[x265] [PATCH] asm: idct[8x8] sse2 12232 -> 3500 over c code 3550 -> 3500 over intrinsic
dave
dtyx265 at gmail.com
Wed Dec 17 03:10:32 CET 2014
Sorry for the slow reply, I ran into some unrelated technical
difficulties...
On 12/09/2014 09:56 AM, chen wrote:
> At 2014-12-09 12:21:22,dtyx265 at gmail.com wrote:
> ># HG changeset patch
> ># User David T Yuendtyx265 at gmail.com>
> ># Date 1418098810 28800
> ># Node ID 39dfcbf07ae468ca9090e2dabb350cc193060229
> ># Parent 53f7efef5ebda6d5ff03e868f2b702c385d72ddd
> >asm: idct[8x8] sse2 12232 -> 3500 over c code 3550 -> 3500 over intrinsic
> >
> <mailto:dtyx265 at gmail.com%3E%3E#%A0Date%A01418098810%A028800%3E#%A0Node%A0ID%A039dfcbf07ae468ca9090e2dabb350cc193060229%3E#%A0Parent%A0%A053f7efef5ebda6d5ff03e868f2b702c385d72ddd%3Easm:%A0idct[8x8]%A0sse2%A012232%A0-%3E%A03500%A0over%A0c%A0code%A0%A03550%A0-%3E%A03500%A0over%A0intrinsic%3E>--- a/source/common/x86/dct8.asm Sat Dec 06 17:17:59 2014 +0900
> >+++ b/source/common/x86/dct8.asm Mon Dec 08 20:20:10 2014 -0800
> >@@ -974,6 +974,386 @@
> > RET
> >
> > ;-------------------------------------------------------
> >+; void idct8(int32_t* src, int16_t* dst, intptr_t dstStride)
> >+;-------------------------------------------------------
> >+INIT_XMM sse2
> >+%if ARCH_X86_64
> >+%if BIT_DEPTH == 10
> >+ %define IDCT_SHIFT 10
> >+ %define IDCT_ADD pd_512
> >+%elif BIT_DEPTH == 8
> >+ %define IDCT_SHIFT 12
> >+ %define IDCT_ADD pd_2048
> >+%else
> >+ %error Unsupported BIT_DEPTH!
> >+%endif
> >+
> >+cglobal idct8, 3,6, 16
> cglobal fourth parameter is stack space
Understood....
>
> >+ mov r5, rsp
> >+ sub rsp, 5 * mmsize
> >+ and rsp, ~(16-1)
> >+ mova m9, [r0 + 1 * mmsize]
> >+ mova m1, [r0 + 3 * mmsize]
> >+ mova m7, m9
> >+ punpcklwd m7, m1
> >+ punpckhwd m9, m1
> >+ mova m14, [tab_idct8_3]
> >+ mova m3, m14
> >+ pmaddwd m14, m7
> >+ pmaddwd m3, m9
> >+ mova m0, [r0 + 5 * mmsize]
> >+ mova m10, [r0 + 7 * mmsize]
> >+ mova m2, m0
> >+ punpcklwd m2, m10
> >+ punpckhwd m0, m10
> >+ mova m15, [tab_idct8_3 + 1 * mmsize]
> >+ mova m11, [tab_idct8_3 + 1 * mmsize]
> >+ pmaddwd m15, m2
> >+ mova m4, [tab_idct8_3 + 2 * mmsize]
> >+ pmaddwd m11, m0
> >+ mova m1, [tab_idct8_3 + 2 * mmsize]
> >+ paddd m15, m14
> >+ mova m5, [tab_idct8_3 + 4 * mmsize]
> >+ mova m12, [tab_idct8_3 + 4 * mmsize]
> >+ paddd m11, m3
> >+ mova [rsp + 0 * mmsize], m11
> >+ mova [rsp + 1 * mmsize], m15
> >+ pmaddwd m4, m7
> >+ pmaddwd m1, m9
> >+ mova m14, [tab_idct8_3 + 3 * mmsize]
> >+ mova m3, [tab_idct8_3 + 3 * mmsize]
> >+ pmaddwd m14, m2
> >+ pmaddwd m3, m0
> >+ paddd m14, m4
> >+ paddd m3, m1
> >+ mova [rsp + 2 * mmsize], m3
> >+ pmaddwd m5, m9
> >+ pmaddwd m9, [tab_idct8_3 + 6 * mmsize]
> >+ mova m6, [tab_idct8_3 + 5 * mmsize]
> >+ pmaddwd m12, m7
> >+ pmaddwd m7, [tab_idct8_3 + 6 * mmsize]
> >+ mova m4, [tab_idct8_3 + 5 * mmsize]
> >+ pmaddwd m6, m2
> >+ paddd m6, m12
> >+ pmaddwd m2, [tab_idct8_3 + 7 * mmsize]
> >+ paddd m7, m2
> we can reduce register number, here is a example, m2 is free now
> >+ mova [rsp + 3 * mmsize], m6
> >+ pmaddwd m4, m0
> >+ pmaddwd m0, [tab_idct8_3 + 7 * mmsize]
> >+ paddd m9, m0
> >+ paddd m5, m4
> >+ mova m6, [r0 + 0 * mmsize]
> m6 can replace by m2
I don't see how to get this to work. m2 and m6 are reused and I don't
see any other free registers to replace them with. I have already made
as many improvements of the type you are suggesting as I could find.
>
> >+ mova m0, [r0 + 4 * mmsize]
> >+ mova m4, m6
> >+ punpcklwd m4, m0
> >+ punpckhwd m6, m0
> >+ mova m12, [r0 + 2 * mmsize]
> >+ mova m0, [r0 + 6 * mmsize]
> >+ mova m13, m12
> >+ mova m8, [tab_dct4]
> >+ punpcklwd m13, m0
> >+ mova m10, [tab_dct4]
> >+ punpckhwd m12, m0
> >+ pmaddwd m8, m4
> >+ mova m3, m8
> >+ pmaddwd m4, [tab_dct4 + 2 * mmsize]
> >+ pmaddwd m10, m6
> >+ mova m2, [tab_dct4 + 1 * mmsize]
> >+ mova m1, m10
> >+ pmaddwd m6, [tab_dct4 + 2 * mmsize]
> >+ mova m0, [tab_dct4 + 1 * mmsize]
> >+ pmaddwd m2, m13
> >+ paddd m3, m2
> >+ psubd m8, m2
> >+ mova m2, m6
> >+ pmaddwd m13, [tab_dct4 + 3 * mmsize]
> >+ pmaddwd m0, m12
> >+ paddd m1, m0
> >+ psubd m10, m0
> >+ mova m0, m4
> >+ pmaddwd m12, [tab_dct4 + 3 * mmsize]
> >+ paddd m3, [pd_64]
> >+ paddd m1, [pd_64]
> >+ paddd m8, [pd_64]
> >+ paddd m10, [pd_64]
> >+ paddd m0, m13
> >+ paddd m2, m12
> >+ paddd m0, [pd_64]
> >+ paddd m2, [pd_64]
> >+ psubd m4, m13
> >+ psubd m6, m12
> >+ paddd m4, [pd_64]
> >+ paddd m6, [pd_64]
> >+ mova m12, m8
> >+ psubd m8, m7
> >+ psrad m8, 7
> >+ paddd m15, m3
> >+ psubd m3, [rsp + 1 * mmsize]
> >+ psrad m15, 7
> >+ paddd m12, m7
> >+ psrad m12, 7
> >+ paddd m11, m1
> >+ mova m13, m14
> >+ psrad m11, 7
> >+ packssdw m15, m11
> >+ psubd m1, [rsp + 0 * mmsize]
> >+ psrad m1, 7
> >+ mova m11, [rsp + 2 * mmsize]
> >+ paddd m14, m0
> >+ psrad m14, 7
> >+ psubd m0, m13
> >+ psrad m0, 7
> >+ paddd m11, m2
> >+ mova m13, [rsp + 3 * mmsize]
> >+ psrad m11, 7
> >+ packssdw m14, m11
> >+ mova m11, m6
> >+ psubd m6, m5
> >+ paddd m13, m4
> >+ psrad m13, 7
> >+ psrad m6, 7
> >+ paddd m11, m5
> >+ psrad m11, 7
> >+ packssdw m13, m11
> >+ mova m11, m10
> >+ psubd m4, [rsp + 3 * mmsize]
> >+ psubd m10, m9
> >+ psrad m4, 7
> >+ psrad m10, 7
> >+ packssdw m4, m6
> >+ packssdw m8, m10
> >+ paddd m11, m9
> >+ psrad m11, 7
> >+ packssdw m12, m11
> >+ psubd m2, [rsp + 2 * mmsize]
> >+ mova m5, m15
> >+ psrad m2, 7
> >+ packssdw m0, m2
> >+ mova m2, m14
> >+ psrad m3, 7
> >+ packssdw m3, m1
> >+ mova m6, m13
> >+ punpcklwd m5, m8
> >+ punpcklwd m2, m4
> >+ mova m1, m12
> >+ punpcklwd m6, m0
> >+ punpcklwd m1, m3
> >+ mova m9, m5
> >+ punpckhwd m13, m0
> >+ mova m0, m2
> >+ punpcklwd m9, m6
> >+ punpckhwd m5, m6
> >+ punpcklwd m0, m1
> >+ punpckhwd m2, m1
> >+ punpckhwd m15, m8
> >+ mova m1, m5
> >+ punpckhwd m14, m4
> >+ punpckhwd m12, m3
> >+ mova m6, m9
> >+ punpckhwd m9, m0
> >+ punpcklwd m1, m2
> >+ mova m4, [tab_idct8_3 + 0 * mmsize]
> >+ punpckhwd m5, m2
> >+ punpcklwd m6, m0
> >+ mova m2, m15
> >+ mova m0, m14
> >+ mova m7, m9
> >+ punpcklwd m2, m13
> >+ punpcklwd m0, m12
> >+ punpcklwd m7, m5
> >+ punpckhwd m14, m12
> >+ mova m10, m2
> >+ punpckhwd m15, m13
> >+ punpckhwd m9, m5
> >+ pmaddwd m4, m7
> >+ mova m13, m1
> >+ punpckhwd m2, m0
> >+ punpcklwd m10, m0
> >+ mova m0, m15
> >+ punpckhwd m15, m14
> >+ mova m12, m1
> >+ mova m3, [tab_idct8_3 + 0 * mmsize]
> >+ punpcklwd m0, m14
> >+ pmaddwd m3, m9
> >+ mova m11, m2
> >+ punpckhwd m2, m15
> >+ punpcklwd m11, m15
> >+ mova m8, [tab_idct8_3 + 1 * mmsize]
> >+ punpcklwd m13, m0
> >+ punpckhwd m12, m0
> >+ pmaddwd m8, m11
> >+ paddd m8, m4
> >+ mova [rsp + 4 * mmsize], m8
> >+ mova m4, [tab_idct8_3 + 2 * mmsize]
> >+ pmaddwd m4, m7
> >+ mova m15, [tab_idct8_3 + 2 * mmsize]
> >+ mova m5, [tab_idct8_3 + 1 * mmsize]
> >+ pmaddwd m15, m9
> >+ pmaddwd m5, m2
> >+ paddd m5, m3
> >+ mova [rsp + 3 * mmsize], m5
> >+ mova m14, [tab_idct8_3 + 3 * mmsize]
> >+ mova m5, [tab_idct8_3 + 3 * mmsize]
> >+ pmaddwd m14, m11
> >+ paddd m14, m4
> >+ mova [rsp + 2 * mmsize], m14
> >+ pmaddwd m5, m2
> >+ paddd m5, m15
> >+ mova [rsp + 1 * mmsize], m5
> >+ mova m15, [tab_idct8_3 + 4 * mmsize]
> >+ mova m5, [tab_idct8_3 + 4 * mmsize]
> >+ pmaddwd m15, m7
> >+ pmaddwd m7, [tab_idct8_3 + 6 * mmsize]
> >+ pmaddwd m5, m9
> >+ pmaddwd m9, [tab_idct8_3 + 6 * mmsize]
> >+ mova m4, [tab_idct8_3 + 5 * mmsize]
> >+ pmaddwd m4, m2
> >+ paddd m5, m4
> >+ mova m4, m6
> >+ mova m8, [tab_idct8_3 + 5 * mmsize]
> >+ punpckhwd m6, m10
> >+ pmaddwd m2, [tab_idct8_3 + 7 * mmsize]
> >+ punpcklwd m4, m10
> >+ paddd m9, m2
> >+ pmaddwd m8, m11
> >+ mova m10, [tab_dct4]
> >+ paddd m8, m15
> >+ pmaddwd m11, [tab_idct8_3 + 7 * mmsize]
> >+ paddd m7, m11
> >+ mova [rsp + 0 * mmsize], m8
> >+ pmaddwd m10, m6
> >+ pmaddwd m6, [tab_dct4 + 2 * mmsize]
> >+ mova m1, m10
> >+ mova m8, [tab_dct4]
> >+ mova m3, [tab_dct4 + 1 * mmsize]
> >+ pmaddwd m8, m4
> >+ pmaddwd m4, [tab_dct4 + 2 * mmsize]
> >+ mova m0, m8
> >+ mova m2, [tab_dct4 + 1 * mmsize]
> >+ pmaddwd m3, m13
> >+ psubd m8, m3
> >+ paddd m0, m3
> >+ mova m3, m6
> >+ pmaddwd m13, [tab_dct4 + 3 * mmsize]
> >+ pmaddwd m2, m12
> >+ paddd m1, m2
> >+ psubd m10, m2
> >+ mova m2, m4
> >+ pmaddwd m12, [tab_dct4 + 3 * mmsize]
> >+ paddd m0, [IDCT_ADD]
> >+ paddd m1, [IDCT_ADD]
> >+ paddd m8, [IDCT_ADD]
> >+ paddd m10, [IDCT_ADD]
> >+ paddd m2, m13
> >+ paddd m3, m12
> >+ paddd m2, [IDCT_ADD]
> >+ paddd m3, [IDCT_ADD]
> >+ psubd m4, m13
> >+ psubd m6, m12
> >+ paddd m4, [IDCT_ADD]
> >+ paddd m6, [IDCT_ADD]
> >+ mova m15, [rsp + 4 * mmsize]
> >+ mova m12, m8
> >+ psubd m8, m7
> >+ psrad m8, IDCT_SHIFT
> >+ mova m11, [rsp + 3 * mmsize]
> >+ paddd m15, m0
> >+ psrad m15, IDCT_SHIFT
> >+ psubd m0, [rsp + 4 * mmsize]
> >+ psrad m0, IDCT_SHIFT
> >+ paddd m12, m7
> >+ paddd m11, m1
> >+ mova m14, [rsp + 2 * mmsize]
> >+ psrad m11, IDCT_SHIFT
> >+ packssdw m15, m11
> >+ psubd m1, [rsp + 3 * mmsize]
> >+ psrad m1, IDCT_SHIFT
> >+ mova m11, [rsp + 1 * mmsize]
> >+ paddd m14, m2
> >+ psrad m14, IDCT_SHIFT
> >+ packssdw m0, m1
> >+ psrad m12, IDCT_SHIFT
> >+ psubd m2, [rsp + 2 * mmsize]
> >+ paddd m11, m3
> >+ mova m13, [rsp + 0 * mmsize]
> >+ psrad m11, IDCT_SHIFT
> >+ packssdw m14, m11
> >+ mova m11, m6
> >+ psubd m6, m5
> >+ paddd m13, m4
> >+ psrad m13, IDCT_SHIFT
> >+ mova m1, m15
> >+ paddd m11, m5
> >+ psrad m11, IDCT_SHIFT
> >+ packssdw m13, m11
> >+ mova m11, m10
> >+ psubd m10, m9
> >+ psrad m10, IDCT_SHIFT
> >+ packssdw m8, m10
> >+ psrad m6, IDCT_SHIFT
> >+ psubd m4, [rsp + 0 * mmsize]
> >+ paddd m11, m9
> >+ psrad m11, IDCT_SHIFT
> >+ packssdw m12, m11
> >+ punpcklwd m1, m14
> >+ mova m5, m13
> >+ psrad m4, IDCT_SHIFT
> >+ packssdw m4, m6
> >+ psubd m3, [rsp + 1 * mmsize]
> >+ psrad m2, IDCT_SHIFT
> >+ mova m6, m8
> >+ psrad m3, IDCT_SHIFT
> >+ punpcklwd m5, m12
> >+ packssdw m2, m3
> >+ punpcklwd m6, m4
> >+ punpckhwd m8, m4
> >+ mova m4, m1
> >+ mova m3, m2
> >+ punpckhdq m1, m5
> >+ punpckldq m4, m5
> >+ punpcklwd m3, m0
> >+ punpckhwd m2, m0
> >+ mova m0, m6
> >+ lea r2, [r2 + r2]
> >+ lea r4, [r2 + r2]
> >+ lea r3, [r4 + r2]
> >+ lea r4, [r4 + r3]
> >+ lea r0, [r4 + r2 * 2]
> >+ movq [r1], m4
> >+ punpckhwd m15, m14
> >+ movhps [r1 + r2], m4
> >+ punpckhdq m0, m3
> >+ movq [r1 + r2 * 2], m1
> >+ punpckhwd m13, m12
> >+ movhps [r1 + r3], m1
> >+ mova m1, m6
> >+ punpckldq m1, m3
> >+ movq [r1 + 8], m1
> >+ movhps [r1 + r2 + 8], m1
> >+ movq [r1 + r2 * 2 + 8], m0
> >+ movhps [r1 + r3 + 8], m0
> >+ mova m0, m15
> >+ punpckhdq m15, m13
> >+ punpckldq m0, m13
> >+ movq [r1 + r2 * 4], m0
> >+ movhps [r1 + r4], m0
> >+ mova m0, m8
> >+ punpckhdq m8, m2
> >+ movq [r1 + r3 * 2], m15
> >+ punpckldq m0, m2
> >+ movhps [r1 + r0], m15
> >+ movq [r1 + r2 * 4 + 8], m0
> >+ movhps [r1 + r4 + 8], m0
> >+ movq [r1 + r3 * 2 + 8], m8
> >+ movhps [r1 + r0 + 8], m8
> >+ mov rsp, r5
> >+ RET
> >+
> >+%undef IDCT_SHIFT
> >+%undef IDCT_ADD
> >+%endif
> >+
> >+;-------------------------------------------------------
> > ; void idct8(const int16_t* src, int16_t* dst, intptr_t dstStride)
> > ;-------------------------------------------------------
> > INIT_XMM ssse3
> >diff -r 53f7efef5ebd -r 39dfcbf07ae4 source/common/x86/dct8.h
> >--- a/source/common/x86/dct8.h Sat Dec 06 17:17:59 2014 +0900
> >+++ b/source/common/x86/dct8.h Mon Dec 08 20:20:10 2014 -0800
> >@@ -34,6 +34,9 @@
> > void x265_idst4_sse2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> > void x265_idct4_sse2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> > void x265_idct4_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> >+#if X86_64
> >+void x265_idct8_sse2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> >+#endif
> > void x265_idct8_ssse3(const int16_t* src, int16_t* dst, intptr_t dstStride);
> > void x265_idct8_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> > void x265_idct16_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
> >_______________________________________________
> >x265-devel mailing list
> >x265-devel at videolan.org
> >https://mailman.videolan.org/listinfo/x265-devel
>
>
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20141216/f0e35805/attachment-0001.html>
More information about the x265-devel
mailing list