[x264-devel] x86: AVX-512 dequant_8x8

[Henrik Gramner]

x264 | branch: master | Henrik Gramner <henrik at gramner.com> | Tue Apr  4 20:54:12 2017 +0200| [40aca29a164d5e5e6589d507bdcae6717d72f6bf] | committer: Henrik Gramner

x86: AVX-512 dequant_8x8

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=40aca29a164d5e5e6589d507bdcae6717d72f6bf
---

 common/quant.c         |  4 +++
 common/x86/quant-a.asm | 79 +++++++++++++++++++++++++++++++++++++++++++++++++-
 common/x86/quant.h     |  1 +
 3 files changed, 83 insertions(+), 1 deletion(-)

diff --git a/common/quant.c b/common/quant.c
index 5710356a..37ae7b92 100644
--- a/common/quant.c
+++ b/common/quant.c
@@ -561,6 +561,7 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
     if( cpu&X264_CPU_AVX512 )
     {
         pf->dequant_4x4 = x264_dequant_4x4_avx512;
+        pf->dequant_8x8 = x264_dequant_8x8_avx512;
         pf->coeff_last4 = x264_coeff_last4_avx512;
         pf->coeff_last8 = x264_coeff_last8_avx512;
         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_avx512;
@@ -729,7 +730,10 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
     if( cpu&X264_CPU_AVX512 )
     {
         if( h->param.i_cqm_preset != X264_CQM_FLAT )
+        {
             pf->dequant_4x4 = x264_dequant_4x4_avx512;
+            pf->dequant_8x8 = x264_dequant_8x8_avx512;
+        }
         pf->coeff_last8 = x264_coeff_last8_avx512;
         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_avx512;
         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_avx512;
diff --git a/common/x86/quant-a.asm b/common/x86/quant-a.asm
index b797e418..0803b6aa 100644
--- a/common/x86/quant-a.asm
+++ b/common/x86/quant-a.asm
@@ -30,7 +30,12 @@
 %include "x86inc.asm"
 %include "x86util.asm"
 
-SECTION_RODATA 32
+SECTION_RODATA 64
+
+%if HIGH_BIT_DEPTH == 0
+dequant_shuf_avx512: dw  0, 2, 4, 6, 8,10,12,14,16,18,20,22,24,26,28,30
+                     dw 32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62
+%endif
 
 %macro DQM4 3
     dw %1, %2, %1, %2, %2, %3, %2, %3
@@ -800,6 +805,78 @@ cglobal dequant_4x4, 0,3
 %endif
     RET
 
+cglobal dequant_8x8, 0,3
+    DEQUANT_START_AVX512 8
+    mova          m0, [dmf+0*64]
+    mova          m1, [dmf+1*64]
+    mova          m2, [dmf+2*64]
+    mova          m3, [dmf+3*64]
+%if HIGH_BIT_DEPTH
+    pmaddwd       m0, [r0+0*64]
+    pmaddwd       m1, [r0+1*64]
+    pmaddwd       m2, [r0+2*64]
+    pmaddwd       m3, [r0+3*64]
+%else
+    mova          m6, [dequant_shuf_avx512]
+%endif
+    sub          t0d, 6
+    jl .rshift
+%if HIGH_BIT_DEPTH
+    vpbroadcastd  m4, t0d
+    vpsllvd       m0, m4
+    vpsllvd       m1, m4
+    vpsllvd       m2, m4
+    vpsllvd       m3, m4
+    jmp .end
+.rshift:
+%else
+    vpbroadcastw  m4, t0d
+    vpermt2w      m0, m6, m1
+    vpermt2w      m2, m6, m3
+    pmullw        m0, [r0]
+    pmullw        m2, [r0+64]
+    vpsllvw       m0, m4
+    vpsllvw       m2, m4
+    mova        [r0], m0
+    mova     [r0+64], m2
+    RET
+.rshift:
+    pmovzxwd      m4, [r0+0*32]
+    pmovzxwd      m5, [r0+1*32]
+    pmaddwd       m0, m4
+    pmaddwd       m1, m5
+    pmovzxwd      m4, [r0+2*32]
+    pmovzxwd      m5, [r0+3*32]
+    pmaddwd       m2, m4
+    pmaddwd       m3, m5
+%endif
+    mov          r1d, 1<<31
+    shrx         r1d, r1d, t0d ; 1 << (-i_qbits-1)
+    neg          t0d
+    vpbroadcastd  m4, r1d
+    vpbroadcastd  m5, t0d
+    paddd         m0, m4
+    paddd         m1, m4
+    vpsravd       m0, m5
+    vpsravd       m1, m5
+    paddd         m2, m4
+    paddd         m3, m4
+    vpsravd       m2, m5
+    vpsravd       m3, m5
+%if HIGH_BIT_DEPTH
+.end:
+    mova   [r0+0*64], m0
+    mova   [r0+1*64], m1
+    mova   [r0+2*64], m2
+    mova   [r0+3*64], m3
+%else
+    vpermt2w      m0, m6, m1
+    vpermt2w      m2, m6, m3
+    mova        [r0], m0
+    mova     [r0+64], m2
+%endif
+    RET
+
 %undef dmf
 
 %macro DEQUANT_DC 2
diff --git a/common/x86/quant.h b/common/x86/quant.h
index 67cc1e6b..7d28670a 100644
--- a/common/x86/quant.h
+++ b/common/x86/quant.h
@@ -67,6 +67,7 @@ void x264_dequant_4x4_avx2( dctcoef dct[16], int dequant_mf[6][16], int i_qp );
 void x264_dequant_4x4dc_avx2( dctcoef dct[16], int dequant_mf[6][16], int i_qp );
 void x264_dequant_8x8_avx2( dctcoef dct[64], int dequant_mf[6][64], int i_qp );
 void x264_dequant_4x4_avx512( dctcoef dct[16], int dequant_mf[6][16], int i_qp );
+void x264_dequant_8x8_avx512( dctcoef dct[64], int dequant_mf[6][64], int i_qp );
 void x264_dequant_4x4_flat16_mmx( int16_t dct[16], int dequant_mf[6][16], int i_qp );
 void x264_dequant_8x8_flat16_mmx( int16_t dct[64], int dequant_mf[6][64], int i_qp );
 void x264_dequant_4x4_flat16_sse2( int16_t dct[16], int dequant_mf[6][16], int i_qp );