[x265] [PATCH 089 of 307] x86: AVX512 interp_8tap_horiz_pp_32xN

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar
# Date 1504241683 -19800
#      Fri Sep 01 10:24:43 2017 +0530
# Node ID 4be3c35eb7510f269a548f248e4f5904b4107d74
# Parent  354f848c3793b459c005667cdf7158eb6394eb0f
x86: AVX512 interp_8tap_horiz_pp_32xN

Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
32x8  |      18.92x       |      37.84x
32x16 |      17.46x       |      36.15x
32x24 |      17.77x       |      35.98x
32x32 |      17.91x       |      36.69x
32x64 |      18.10x       |      35.47x

diff -r 354f848c3793 -r 4be3c35eb751 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Fri Aug 11 17:18:16 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Fri Sep 01 10:24:43 2017 +0530
@@ -4052,6 +4052,12 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hps = PFX(interp_4tap_horiz_ps_32x24_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_hps = PFX(interp_4tap_horiz_ps_32x8_avx512);
+
+        p.pu[LUMA_32x8].luma_hpp = PFX(interp_8tap_horiz_pp_32x8_avx512);
+        p.pu[LUMA_32x16].luma_hpp = PFX(interp_8tap_horiz_pp_32x16_avx512);
+        p.pu[LUMA_32x24].luma_hpp = PFX(interp_8tap_horiz_pp_32x24_avx512);
+        p.pu[LUMA_32x32].luma_hpp = PFX(interp_8tap_horiz_pp_32x32_avx512);
+        p.pu[LUMA_32x64].luma_hpp = PFX(interp_8tap_horiz_pp_32x64_avx512);
         p.pu[LUMA_64x16].luma_hpp = PFX(interp_8tap_horiz_pp_64x16_avx512);
         p.pu[LUMA_64x32].luma_hpp = PFX(interp_8tap_horiz_pp_64x32_avx512);
         p.pu[LUMA_64x48].luma_hpp = PFX(interp_8tap_horiz_pp_64x48_avx512);
diff -r 354f848c3793 -r 4be3c35eb751 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Fri Aug 11 17:18:16 2017 +0530
+++ b/source/common/x86/ipfilter8.asm	Fri Sep 01 10:24:43 2017 +0530
@@ -10182,6 +10182,57 @@
     movu              [r2],     m7
 %endmacro
 
+%macro PROCESS_IPFILTER_LUMA_PP_32x2_AVX512 0
+    ; register map
+    ; m0 , m1 interpolate coeff
+    ; m2 , m3, m4  shuffle order table
+    ; m5 - pw_1
+    ; m6 - pw_512
+
+    movu             ym7,        [r0]
+    vinserti32x8      m7,        [r0 + r1], 1
+    movu             ym9,        [r0 + 8]
+    vinserti32x8      m9,        [r0 + r1 + 8], 1
+
+    pshufb            m8,        m7,        m3
+    pshufb            m7,        m2
+    pshufb            m10,       m9,        m3
+    pshufb            m11,       m9,        m4
+    pshufb            m9,        m2
+
+    pmaddubsw         m7,        m0
+    pmaddubsw         m12,       m8,        m1
+    pmaddwd           m7,        m5
+    pmaddwd           m12,       m5
+    paddd             m7,        m12
+
+    pmaddubsw         m8,        m0
+    pmaddubsw         m12,       m9,        m1
+    pmaddwd           m8,        m5
+    pmaddwd           m12,       m5
+    paddd             m8,        m12
+
+    pmaddubsw         m9,        m0
+    pmaddubsw         m12,       m10,       m1
+    pmaddwd           m9,        m5
+    pmaddwd           m12,       m5
+    paddd             m9,        m12
+
+    pmaddubsw         m10,       m0
+    pmaddubsw         m12,      m11,        m1
+    pmaddwd           m10,      m5
+    pmaddwd           m12,      m5
+    paddd             m10,      m12
+
+    packssdw          m7,       m8
+    packssdw          m9,       m10
+    pmulhrsw          m7,       m6
+    pmulhrsw          m9,       m6
+    packuswb          m7,       m9
+    movu              [r2],     ym7
+    vextracti32x8     [r2 + r3], m7, 1
+%endmacro
+
 %macro IPFILTER_LUMA_64xN_AVX512 1
 INIT_ZMM avx512
 cglobal interp_8tap_horiz_pp_64x%1, 4,6,13
@@ -10214,6 +10265,40 @@
 IPFILTER_LUMA_64xN_AVX512 32
 IPFILTER_LUMA_64xN_AVX512 48
 IPFILTER_LUMA_64xN_AVX512 64
+
+%macro IPFILTER_LUMA_32xN_AVX512 1
+INIT_ZMM avx512
+cglobal interp_8tap_horiz_pp_32x%1, 4,6,13
+    sub               r0,    3
+    mov               r4d,   r4m
+%ifdef PIC
+    lea               r5,        [tab_LumaCoeff]
+    vpbroadcastd      m0,        [r5 + r4 * 8]
+    vpbroadcastd      m1,        [r5 + r4 * 8 + 4]
+%else
+    vpbroadcastd      m0,        [tab_LumaCoeff + r4 * 8]
+    vpbroadcastd      m1,        [tab_LumaCoeff + r4 * 8 + 4]
+%endif
+    vbroadcasti32x8   m2,        [interp4_horiz_shuf_load1_avx512]
+    vbroadcasti32x8   m3,        [interp4_horiz_shuf_load3_avx512]
+    vbroadcasti32x8   m4,        [interp4_horiz_shuf_load2_avx512]
+    vpbroadcastd      m5,        [pw_1]
+    vbroadcasti32x8   m6,        [pw_512]
+
+%rep %1/2 -1
+    PROCESS_IPFILTER_LUMA_PP_32x2_AVX512
+    lea               r0,        [r0 + 2 * r1]
+    lea               r2,        [r2 + 2 * r3]
+%endrep
+    PROCESS_IPFILTER_LUMA_PP_32x2_AVX512
+    RET
+%endmacro
+
+IPFILTER_LUMA_32xN_AVX512 8
+IPFILTER_LUMA_32xN_AVX512 16
+IPFILTER_LUMA_32xN_AVX512 24
+IPFILTER_LUMA_32xN_AVX512 32
+IPFILTER_LUMA_32xN_AVX512 64
 ;-------------------------------------------------------------------------------------------------------------
 ;ipfilter_luma_avx512 code end
 ;-------------------------------------------------------------------------------------------------------------
\ No newline at end of file