[x265] [PATCH] asm: psyCost_pp_16x16 in sse4: improve 27086c->3566c

Divya Manivannan divya at multicorewareinc.com
Wed Dec 31 08:26:23 CET 2014


# HG changeset patch
# User Divya Manivannan <divya at multicorewareinc.com>
# Date 1420010754 -19800
#      Wed Dec 31 12:55:54 2014 +0530
# Node ID e71e8732ddbc9356ed2e80fcba9472e9188bc735
# Parent  b79da00cd8ad8a2373d6d9c8db2490898106b18d
asm: psyCost_pp_16x16 in sse4: improve 27086c->3566c

diff -r b79da00cd8ad -r e71e8732ddbc source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Wed Dec 31 11:24:09 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Dec 31 12:55:54 2014 +0530
@@ -1438,6 +1438,7 @@
         p.psy_cost_pp[BLOCK_4x4] = x265_psyCost_pp_4x4_sse4;
 #if X86_64
         p.psy_cost_pp[BLOCK_8x8] = x265_psyCost_pp_8x8_sse4;
+        p.psy_cost_pp[BLOCK_16x16] = x265_psyCost_pp_16x16_sse4;
 #endif
     }
     if (cpuMask & X265_CPU_XOP)
@@ -1723,6 +1724,7 @@
         p.psy_cost_pp[BLOCK_4x4] = x265_psyCost_pp_4x4_sse4;
 #if X86_64
         p.psy_cost_pp[BLOCK_8x8] = x265_psyCost_pp_8x8_sse4;
+        p.psy_cost_pp[BLOCK_16x16] = x265_psyCost_pp_16x16_sse4;
 #endif
     }
     if (cpuMask & X265_CPU_AVX)
diff -r b79da00cd8ad -r e71e8732ddbc source/common/x86/pixel-a.asm
--- a/source/common/x86/pixel-a.asm	Wed Dec 31 11:24:09 2014 +0530
+++ b/source/common/x86/pixel-a.asm	Wed Dec 31 12:55:54 2014 +0530
@@ -6920,3 +6920,214 @@
 %endif ; HIGH_BIT_DEPTH
     RET
 %endif
+
+%if ARCH_X86_64
+%if HIGH_BIT_DEPTH
+INIT_XMM sse4
+cglobal psyCost_pp_16x16, 4, 9, 14
+
+    FIX_STRIDES r1, r3
+    lea             r4, [3 * r1]
+    lea             r8, [3 * r3]
+    mova            m12, [pw_1]
+    mova            m13, [pd_1]
+    pxor            m11, m11
+    mov             r7d, 2
+.loopH:
+    mov             r6d, 2
+.loopW:
+    pxor            m10, m10
+    movu            m0, [r0]
+    movu            m1, [r0 + r1]
+    movu            m2, [r0 + r1 * 2]
+    movu            m3, [r0 + r4]
+    lea             r5, [r0 + r1 * 4]
+    movu            m4, [r5]
+    movu            m5, [r5 + r1]
+    movu            m6, [r5 + r1 * 2]
+    movu            m7, [r5 + r4]
+
+    paddw           m8, m0, m1
+    paddw           m8, m2
+    paddw           m8, m3
+    paddw           m8, m4
+    paddw           m8, m5
+    paddw           m8, m6
+    paddw           m8, m7
+    pmaddwd         m8, m12
+    movhlps         m9, m8
+    paddd           m8, m9
+    psrldq          m9, m8, 4
+    paddd           m8, m9
+    psrld           m8, 2
+
+    HADAMARD8_2D 0, 1, 2, 3, 4, 5, 6, 7, 9, amax
+
+    paddd           m0, m1
+    paddd           m0, m2
+    paddd           m0, m3
+    HADDUW m0, m1
+    paddd           m0, m13
+    psrld           m0, 1
+    psubd           m10, m0, m8
+
+    movu            m0, [r2]
+    movu            m1, [r2 + r3]
+    movu            m2, [r2 + r3 * 2]
+    movu            m3, [r2 + r8]
+    lea             r5, [r2 + r3 * 4]
+    movu            m4, [r5]
+    movu            m5, [r5 + r3]
+    movu            m6, [r5 + r3 * 2]
+    movu            m7, [r5 + r8]
+
+    paddw           m8, m0, m1
+    paddw           m8, m2
+    paddw           m8, m3
+    paddw           m8, m4
+    paddw           m8, m5
+    paddw           m8, m6
+    paddw           m8, m7
+    pmaddwd         m8, m12
+    movhlps         m9, m8
+    paddd           m8, m9
+    psrldq          m9, m8, 4
+    paddd           m8, m9
+    psrld           m8, 2
+
+    HADAMARD8_2D 0, 1, 2, 3, 4, 5, 6, 7, 9, amax
+
+    paddd           m0, m1
+    paddd           m0, m2
+    paddd           m0, m3
+    HADDUW m0, m1
+    paddd           m0, m13
+    psrld           m0, 1
+    psubd           m0, m8
+    psubd           m10, m0
+    pabsd           m0, m10
+    paddd           m11, m0
+    add             r0, 16
+    add             r2, 16
+    dec             r6d
+    jnz             .loopW
+    lea             r0, [r0 + r1 * 8 - 32]
+    lea             r2, [r2 + r3 * 8 - 32]
+    dec             r7d
+    jnz             .loopH
+    movd            eax, m11
+    RET
+%else ; !HIGH_BIT_DEPTH
+INIT_XMM sse4
+cglobal psyCost_pp_16x16, 4, 9, 15
+    lea             r4, [3 * r1]
+    lea             r8, [3 * r3]
+    mova            m8, [hmul_8p]
+    mova            m10, [pw_1]
+    mova            m14, [pd_1]
+    pxor            m13, m13
+    mov             r7d, 2
+.loopH:
+    mov             r6d, 2
+.loopW:
+    pxor            m12, m12
+    movddup         m0, [r0]
+    movddup         m1, [r0 + r1]
+    movddup         m2, [r0 + r1 * 2]
+    movddup         m3, [r0 + r4]
+    lea             r5, [r0 + r1 * 4]
+    movddup         m4, [r5]
+    movddup         m5, [r5 + r1]
+    movddup         m6, [r5 + r1 * 2]
+    movddup         m7, [r5 + r4]
+
+    pmaddubsw       m0, m8
+    pmaddubsw       m1, m8
+    pmaddubsw       m2, m8
+    pmaddubsw       m3, m8
+    pmaddubsw       m4, m8
+    pmaddubsw       m5, m8
+    pmaddubsw       m6, m8
+    pmaddubsw       m7, m8
+
+    paddw           m11, m0, m1
+    paddw           m11, m2
+    paddw           m11, m3
+    paddw           m11, m4
+    paddw           m11, m5
+    paddw           m11, m6
+    paddw           m11, m7
+
+    pmaddwd         m11, m10
+    psrldq          m9, m11, 4
+    paddd           m11, m9
+    psrld           m11, 2
+
+    HADAMARD8_2D_HMUL 0, 1, 2, 3, 4, 5, 6, 7, 9, 9
+
+    paddw           m0, m1
+    paddw           m0, m2
+    paddw           m0, m3
+    HADDW m0, m1
+
+    paddd           m0, m14
+    psrld           m0, 1
+    psubd           m12, m0, m11
+
+    movddup         m0, [r2]
+    movddup         m1, [r2 + r3]
+    movddup         m2, [r2 + r3 * 2]
+    movddup         m3, [r2 + r8]
+    lea             r5, [r2 + r3 * 4]
+    movddup         m4, [r5]
+    movddup         m5, [r5 + r3]
+    movddup         m6, [r5 + r3 * 2]
+    movddup         m7, [r5 + r8]
+
+    pmaddubsw       m0, m8
+    pmaddubsw       m1, m8
+    pmaddubsw       m2, m8
+    pmaddubsw       m3, m8
+    pmaddubsw       m4, m8
+    pmaddubsw       m5, m8
+    pmaddubsw       m6, m8
+    pmaddubsw       m7, m8
+
+    paddw           m11, m0, m1
+    paddw           m11, m2
+    paddw           m11, m3
+    paddw           m11, m4
+    paddw           m11, m5
+    paddw           m11, m6
+    paddw           m11, m7
+
+    pmaddwd         m11, m10
+    psrldq          m9, m11, 4
+    paddd           m11, m9
+    psrld           m11, 2
+
+    HADAMARD8_2D_HMUL 0, 1, 2, 3, 4, 5, 6, 7, 9, 9
+
+    paddw           m0, m1
+    paddw           m0, m2
+    paddw           m0, m3
+    HADDW m0, m1
+
+    paddd           m0, m14
+    psrld           m0, 1
+    psubd           m0, m11
+    psubd           m12, m0
+    pabsd           m0, m12
+    paddd           m13, m0
+    add             r0, 8
+    add             r2, 8
+    dec             r6d
+    jnz             .loopW
+    lea             r0, [r0 + r1 * 8 - 16]
+    lea             r2, [r2 + r3 * 8 - 16]
+    dec             r7d
+    jnz             .loopH
+    movd            eax, m13
+    RET
+%endif ; HIGH_BIT_DEPTH
+%endif
diff -r b79da00cd8ad -r e71e8732ddbc source/common/x86/pixel.h
--- a/source/common/x86/pixel.h	Wed Dec 31 11:24:09 2014 +0530
+++ b/source/common/x86/pixel.h	Wed Dec 31 12:55:54 2014 +0530
@@ -220,6 +220,7 @@
 void x265_upShift_8_sse4(const uint8_t* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int width, int height, int shift);
 int x265_psyCost_pp_4x4_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
 int x265_psyCost_pp_8x8_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
+int x265_psyCost_pp_16x16_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
 
 #undef DECL_PIXELS
 #undef DECL_HEVC_SSD


More information about the x265-devel mailing list