[x264-devel] [PATCH 9/9] MIPS MSA Quant module optimization
Rishikesh More
rishikesh.more at imgtec.com
Thu Jun 18 14:18:46 CEST 2015
This patch adds MSA (MIPS-SIMD-Arch) optimizations for quant functions.
Signed-off-by: Rishikesh More <rishikesh.more at imgtec.com>
---
Makefile | 2 +-
common/mips/quant-c.c | 630 ++++++++++++++++++++++++++++++++++++++++++++++++++
common/mips/quant.h | 43 ++++
common/quant.c | 18 ++
4 files changed, 692 insertions(+), 1 deletion(-)
create mode 100644 common/mips/quant-c.c
create mode 100644 common/mips/quant.h
diff --git a/Makefile b/Makefile
index fd32b9f..6193c59 100644
--- a/Makefile
+++ b/Makefile
@@ -148,7 +148,7 @@ ifeq ($(SYS_ARCH),MIPS)
ifneq ($(findstring HAVE_MSA 1, $(CONFIG)),)
SRCS += common/mips/mc-c.c common/mips/dct-c.c \
common/mips/deblock-c.c common/mips/pixel-c.c \
- common/mips/predict-c.c
+ common/mips/predict-c.c common/mips/quant-c.c
endif
endif
diff --git a/common/mips/quant-c.c b/common/mips/quant-c.c
new file mode 100644
index 0000000..8deddc6
--- /dev/null
+++ b/common/mips/quant-c.c
@@ -0,0 +1,630 @@
+/*****************************************************************************
+ * quant-c.c: msa quantization and level-run
+ *****************************************************************************
+ * Copyright (C) 2015 x264 project
+ *
+ * Authors: Rishikesh More <rishikesh.more at imgtec.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
+ *
+ * This program is also available under a commercial proprietary license.
+ * For more information, contact us at licensing at x264.com.
+ *****************************************************************************/
+
+#include "common/common.h"
+#include "macros.h"
+
+#if !HIGH_BIT_DEPTH
+static void avc_dequant_4x4_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][16],
+ int32_t i_qp )
+{
+ const int32_t i_mf = i_qp % 6;
+ const int32_t q_bits = i_qp / 6 - 4;
+ v8i16 dct0, dct1;
+ v4i32 dequant_m_f0, dequant_m_f1, dequant_m_f2, dequant_m_f3;
+
+ LD_SH2( p_dct, 8, dct0, dct1 );
+
+ LD_SW2( pi_dequant_mf[i_mf], 4, dequant_m_f0, dequant_m_f1 );
+ LD_SW2( pi_dequant_mf[i_mf] + 8, 4, dequant_m_f2, dequant_m_f3 );
+
+ if ( q_bits >= 0 )
+ {
+ v8i16 dequant_mf_h0, dequant_mf_h1, q_bits_vec;
+
+ q_bits_vec = __msa_fill_h( q_bits );
+
+ PCKEV_H2_SH( dequant_m_f1, dequant_m_f0, dequant_m_f3, dequant_m_f2,
+ dequant_mf_h0, dequant_mf_h1 );
+
+ dct0 *= dequant_mf_h0;
+ dct1 *= dequant_mf_h1;
+ dct0 <<= q_bits_vec;
+ dct1 <<= q_bits_vec;
+ ST_SH2( dct0, dct1, p_dct, 8 );
+ }
+ else
+ {
+ const int32_t q_bits_add = 1 << ( -q_bits - 1 );
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+ v4i32 q_bits_vec, q_bits_vec_add;
+
+ q_bits_vec_add = __msa_fill_w( q_bits_add );
+ q_bits_vec = __msa_fill_w( -q_bits );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+
+ dct_signed_w0 *= dequant_m_f0;
+ dct_signed_w1 *= dequant_m_f1;
+ dct_signed_w2 *= dequant_m_f2;
+ dct_signed_w3 *= dequant_m_f3;
+ dct_signed_w0 += q_bits_vec_add;
+ dct_signed_w1 += q_bits_vec_add;
+ dct_signed_w2 += q_bits_vec_add;
+ dct_signed_w3 += q_bits_vec_add;
+
+ SRA_4V( dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3,
+ q_bits_vec );
+ PCKEV_H2_SH( dct_signed_w1, dct_signed_w0, dct_signed_w3, dct_signed_w2,
+ dct0, dct1 );
+ ST_SH2( dct0, dct1, p_dct, 8 );
+ }
+}
+
+static void avc_dequant_8x8_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][64],
+ int32_t i_qp )
+{
+ const int32_t i_mf = i_qp % 6;
+ const int32_t q_bits = i_qp / 6 - 6;
+ v8i16 dct0, dct1, dct2, dct3, dct4, dct5, dct6, dct7;
+ v4i32 dequant_m_f0, dequant_m_f1, dequant_m_f2, dequant_m_f3;
+ v4i32 dequant_m_f4, dequant_m_f5, dequant_m_f6, dequant_m_f7;
+ v4i32 dequant_m_f8, dequant_m_f9, dequant_m_f10, dequant_m_f11;
+ v4i32 dequant_m_f12, dequant_m_f13, dequant_m_f14, dequant_m_f15;
+
+ LD_SH8( p_dct, 8, dct0, dct1, dct2, dct3, dct4, dct5, dct6, dct7 );
+
+ LD_SW2( pi_dequant_mf[i_mf], 4, dequant_m_f0, dequant_m_f1 );
+ LD_SW2( pi_dequant_mf[i_mf] + 8, 4, dequant_m_f2, dequant_m_f3 );
+ LD_SW2( pi_dequant_mf[i_mf] + 16, 4, dequant_m_f4, dequant_m_f5 );
+ LD_SW2( pi_dequant_mf[i_mf] + 24, 4, dequant_m_f6, dequant_m_f7 );
+ LD_SW2( pi_dequant_mf[i_mf] + 32, 4, dequant_m_f8, dequant_m_f9 );
+ LD_SW2( pi_dequant_mf[i_mf] + 40, 4, dequant_m_f10, dequant_m_f11 );
+ LD_SW2( pi_dequant_mf[i_mf] + 48, 4, dequant_m_f12, dequant_m_f13 );
+ LD_SW2( pi_dequant_mf[i_mf] + 56, 4, dequant_m_f14, dequant_m_f15 );
+
+ if ( q_bits >= 0 )
+ {
+ v8i16 q_bits_vec;
+ v8i16 dequant_mf_h0, dequant_mf_h1, dequant_mf_h2, dequant_mf_h3;
+ v8i16 dequant_mf_h4, dequant_mf_h5, dequant_mf_h6, dequant_mf_h7;
+
+ q_bits_vec = __msa_fill_h( q_bits );
+
+ PCKEV_H4_SH( dequant_m_f1, dequant_m_f0, dequant_m_f3, dequant_m_f2,
+ dequant_m_f5, dequant_m_f4, dequant_m_f7, dequant_m_f6,
+ dequant_mf_h0, dequant_mf_h1,
+ dequant_mf_h2, dequant_mf_h3 );
+ PCKEV_H4_SH( dequant_m_f9, dequant_m_f8, dequant_m_f11, dequant_m_f10,
+ dequant_m_f13, dequant_m_f12, dequant_m_f15, dequant_m_f14,
+ dequant_mf_h4, dequant_mf_h5,
+ dequant_mf_h6, dequant_mf_h7 );
+
+ dct0 *= dequant_mf_h0;
+ dct1 *= dequant_mf_h1;
+ dct2 *= dequant_mf_h2;
+ dct3 *= dequant_mf_h3;
+ dct4 *= dequant_mf_h4;
+ dct5 *= dequant_mf_h5;
+ dct6 *= dequant_mf_h6;
+ dct7 *= dequant_mf_h7;
+
+ SLLI_4V( dct0, dct1, dct2, dct3, q_bits_vec );
+ SLLI_4V( dct4, dct5, dct6, dct7, q_bits_vec );
+
+ ST_SH8( dct0, dct1, dct2, dct3, dct4, dct5, dct6, dct7, p_dct, 8 );
+ }
+ else
+ {
+ const int32_t q_bits_add = 1 << ( -q_bits - 1 );
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+ v4i32 dct_signed_w4, dct_signed_w5, dct_signed_w6, dct_signed_w7;
+ v4i32 dct_signed_w8, dct_signed_w9, dct_signed_w10, dct_signed_w11;
+ v4i32 dct_signed_w12, dct_signed_w13, dct_signed_w14, dct_signed_w15;
+ v4i32 q_bits_vec, q_bits_vec_add;
+
+ q_bits_vec_add = __msa_fill_w( q_bits_add );
+ q_bits_vec = __msa_fill_w( -q_bits );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+ UNPCK_SH_SW( dct2, dct_signed_w4, dct_signed_w5 );
+ UNPCK_SH_SW( dct3, dct_signed_w6, dct_signed_w7 );
+ UNPCK_SH_SW( dct4, dct_signed_w8, dct_signed_w9 );
+ UNPCK_SH_SW( dct5, dct_signed_w10, dct_signed_w11 );
+ UNPCK_SH_SW( dct6, dct_signed_w12, dct_signed_w13 );
+ UNPCK_SH_SW( dct7, dct_signed_w14, dct_signed_w15 );
+
+ dct_signed_w0 *= dequant_m_f0;
+ dct_signed_w1 *= dequant_m_f1;
+ dct_signed_w2 *= dequant_m_f2;
+ dct_signed_w3 *= dequant_m_f3;
+ dct_signed_w4 *= dequant_m_f4;
+ dct_signed_w5 *= dequant_m_f5;
+ dct_signed_w6 *= dequant_m_f6;
+ dct_signed_w7 *= dequant_m_f7;
+ dct_signed_w8 *= dequant_m_f8;
+ dct_signed_w9 *= dequant_m_f9;
+ dct_signed_w10 *= dequant_m_f10;
+ dct_signed_w11 *= dequant_m_f11;
+ dct_signed_w12 *= dequant_m_f12;
+ dct_signed_w13 *= dequant_m_f13;
+ dct_signed_w14 *= dequant_m_f14;
+ dct_signed_w15 *= dequant_m_f15;
+
+ dct_signed_w0 += q_bits_vec_add;
+ dct_signed_w1 += q_bits_vec_add;
+ dct_signed_w2 += q_bits_vec_add;
+ dct_signed_w3 += q_bits_vec_add;
+ dct_signed_w4 += q_bits_vec_add;
+ dct_signed_w5 += q_bits_vec_add;
+ dct_signed_w6 += q_bits_vec_add;
+ dct_signed_w7 += q_bits_vec_add;
+ dct_signed_w8 += q_bits_vec_add;
+ dct_signed_w9 += q_bits_vec_add;
+ dct_signed_w10 += q_bits_vec_add;
+ dct_signed_w11 += q_bits_vec_add;
+ dct_signed_w12 += q_bits_vec_add;
+ dct_signed_w13 += q_bits_vec_add;
+ dct_signed_w14 += q_bits_vec_add;
+ dct_signed_w15 += q_bits_vec_add;
+
+ SRA_4V( dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3,
+ q_bits_vec );
+ SRA_4V( dct_signed_w4, dct_signed_w5, dct_signed_w6, dct_signed_w7,
+ q_bits_vec );
+ SRA_4V( dct_signed_w8, dct_signed_w9, dct_signed_w10, dct_signed_w11,
+ q_bits_vec );
+ SRA_4V( dct_signed_w12, dct_signed_w13, dct_signed_w14, dct_signed_w15,
+ q_bits_vec );
+ PCKEV_H4_SH( dct_signed_w1, dct_signed_w0, dct_signed_w3, dct_signed_w2,
+ dct_signed_w5, dct_signed_w4, dct_signed_w7, dct_signed_w6,
+ dct0, dct1, dct2, dct3 );
+ PCKEV_H4_SH( dct_signed_w9, dct_signed_w8, dct_signed_w11,
+ dct_signed_w10, dct_signed_w13, dct_signed_w12,
+ dct_signed_w15, dct_signed_w14, dct4, dct5, dct6, dct7 );
+ ST_SH8( dct0, dct1, dct2, dct3, dct4, dct5, dct6, dct7, p_dct, 8 );
+ }
+}
+
+static void avc_dequant_4x4_dc_msa( int16_t *p_dct,
+ int32_t pi_dequant_mf[6][16],
+ int32_t i_qp )
+{
+ const int32_t q_bits = i_qp / 6 - 6;
+ int32_t i_dmf = pi_dequant_mf[i_qp % 6][0];
+ v8i16 dct0, dct1, dequant_mf_h;
+
+ LD_SH2( p_dct, 8, dct0, dct1 );
+
+ if ( q_bits >= 0 )
+ {
+ i_dmf <<= q_bits;
+
+ dequant_mf_h = __msa_fill_h( i_dmf );
+ dct0 = dct0 * dequant_mf_h;
+ dct1 = dct1 * dequant_mf_h;
+
+ ST_SH2( dct0, dct1, p_dct, 8 );
+ }
+ else
+ {
+ const int32_t q_bits_add = 1 << ( -q_bits - 1 );
+ v4i32 dequant_m_f, q_bits_vec, q_bits_vec_add;
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+
+ q_bits_vec_add = __msa_fill_w( q_bits_add );
+ q_bits_vec = __msa_fill_w( -q_bits );
+
+ dequant_m_f = __msa_fill_w( i_dmf );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+
+ dct_signed_w0 *= dequant_m_f;
+ dct_signed_w1 *= dequant_m_f;
+ dct_signed_w2 *= dequant_m_f;
+ dct_signed_w3 *= dequant_m_f;
+
+ dct_signed_w0 += q_bits_vec_add;
+ dct_signed_w1 += q_bits_vec_add;
+ dct_signed_w2 += q_bits_vec_add;
+ dct_signed_w3 += q_bits_vec_add;
+
+ SRA_4V( dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3,
+ q_bits_vec );
+ PCKEV_H2_SH( dct_signed_w1, dct_signed_w0, dct_signed_w3, dct_signed_w2,
+ dct0, dct1 );
+ ST_SH2( dct0, dct1, p_dct, 8 );
+ }
+}
+
+static int32_t avc_quant_4x4_msa( int16_t *p_dct, uint16_t *p_mf,
+ uint16_t *p_bias )
+{
+ int32_t non_zero = 0;
+ v8i16 dct0, dct1;
+ v8i16 zero = { 0 };
+ v8i16 dct0_mask, dct1_mask;
+ v8i16 dct_h0, dct_h1, mf_h0, mf_h1, bias_h0, bias_h1;
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+ v4i32 dct_w0, dct_w1, dct_w2, dct_w3;
+ v4i32 mf_vec0, mf_vec1, mf_vec2, mf_vec3;
+ v4i32 bias0, bias1, bias2, bias3;
+
+ LD_SH2( p_dct, 8, dct0, dct1 );
+ LD_SH2( p_bias, 8, bias_h0, bias_h1 );
+ LD_SH2( p_mf, 8, mf_h0, mf_h1 );
+
+ dct0_mask = __msa_clei_s_h( dct0, 0 );
+ dct1_mask = __msa_clei_s_h( dct1, 0 );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+ ILVR_H2_SW( zero, bias_h0, zero, bias_h1, bias0, bias2 );
+ ILVL_H2_SW( zero, bias_h0, zero, bias_h1, bias1, bias3 );
+ ILVR_H2_SW( zero, mf_h0, zero, mf_h1, mf_vec0, mf_vec2 );
+ ILVL_H2_SW( zero, mf_h0, zero, mf_h1, mf_vec1, mf_vec3 );
+
+ dct_w1 = __msa_add_a_w( dct_signed_w1, bias1 );
+ dct_w0 = __msa_add_a_w( dct_signed_w0, bias0 );
+ dct_w2 = __msa_add_a_w( dct_signed_w2, bias2 );
+ dct_w3 = __msa_add_a_w( dct_signed_w3, bias3 );
+
+ dct_w0 *= mf_vec0;
+ dct_w1 *= mf_vec1;
+ dct_w2 *= mf_vec2;
+ dct_w3 *= mf_vec3;
+
+ SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
+ PCKEV_H2_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_h0, dct_h1 );
+
+ dct0 = zero - dct_h0;
+ dct1 = zero - dct_h1;
+
+ dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0, ( v16u8 ) dct0,
+ ( v16u8 ) dct0_mask );
+ dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1, ( v16u8 ) dct1,
+ ( v16u8 ) dct1_mask );
+ non_zero = HADD_SW_S32( ( v4u32 ) ( dct_h0 + dct_h1 ) );
+ ST_SH2( dct0, dct1, p_dct, 8 );
+
+ return !!non_zero;
+}
+
+static int32_t avc_quant_8x8_msa( int16_t *p_dct, uint16_t *p_mf,
+ uint16_t *p_bias )
+{
+ int32_t non_zero = 0;
+ v8i16 dct0, dct1, dct2, dct3;
+ v8i16 zero = { 0 };
+ v8i16 dct0_mask, dct1_mask, dct2_mask, dct3_mask;
+ v8i16 dct_h0, dct_h1, dct_h2, dct_h3, mf_h0, mf_h1, mf_h2, mf_h3;
+ v8i16 bias_h0, bias_h1, bias_h2, bias_h3;
+ v4i32 dct_w0, dct_w1, dct_w2, dct_w3, dct_w4, dct_w5, dct_w6, dct_w7;
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+ v4i32 dct_signed_w4, dct_signed_w5, dct_signed_w6, dct_signed_w7;
+ v4i32 mf_vec0, mf_vec1, mf_vec2, mf_vec3;
+ v4i32 mf_vec4, mf_vec5, mf_vec6, mf_vec7;
+ v4i32 bias0, bias1, bias2, bias3, bias4, bias5, bias6, bias7;
+
+ LD_SH4( p_dct, 8, dct0, dct1, dct2, dct3 );
+
+ dct0_mask = __msa_clei_s_h( dct0, 0 );
+ dct1_mask = __msa_clei_s_h( dct1, 0 );
+ dct2_mask = __msa_clei_s_h( dct2, 0 );
+ dct3_mask = __msa_clei_s_h( dct3, 0 );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+ UNPCK_SH_SW( dct2, dct_signed_w4, dct_signed_w5 );
+ UNPCK_SH_SW( dct3, dct_signed_w6, dct_signed_w7 );
+ LD_SH4( p_bias, 8, bias_h0, bias_h1, bias_h2, bias_h3 );
+ ILVR_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
+ bias0, bias2, bias4, bias6 );
+ ILVL_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
+ bias1, bias3, bias5, bias7 );
+ LD_SH4( p_mf, 8, mf_h0, mf_h1, mf_h2, mf_h3 );
+ ILVR_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
+ mf_vec0, mf_vec2, mf_vec4, mf_vec6 );
+ ILVL_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
+ mf_vec1, mf_vec3, mf_vec5, mf_vec7 );
+
+ dct_w0 = __msa_add_a_w( dct_signed_w0, bias0 );
+ dct_w1 = __msa_add_a_w( dct_signed_w1, bias1 );
+ dct_w2 = __msa_add_a_w( dct_signed_w2, bias2 );
+ dct_w3 = __msa_add_a_w( dct_signed_w3, bias3 );
+ dct_w4 = __msa_add_a_w( dct_signed_w4, bias4 );
+ dct_w5 = __msa_add_a_w( dct_signed_w5, bias5 );
+ dct_w6 = __msa_add_a_w( dct_signed_w6, bias6 );
+ dct_w7 = __msa_add_a_w( dct_signed_w7, bias7 );
+
+ dct_w0 *= mf_vec0;
+ dct_w1 *= mf_vec1;
+ dct_w2 *= mf_vec2;
+ dct_w3 *= mf_vec3;
+ dct_w4 *= mf_vec4;
+ dct_w5 *= mf_vec5;
+ dct_w6 *= mf_vec6;
+ dct_w7 *= mf_vec7;
+
+ SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
+ SRA_4V( dct_w4, dct_w5, dct_w6, dct_w7, 16 );
+ PCKEV_H4_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_w5, dct_w4, dct_w7, dct_w6,
+ dct_h0, dct_h1, dct_h2, dct_h3 );
+ SUB4( zero, dct_h0, zero, dct_h1, zero, dct_h2, zero, dct_h3,
+ dct0, dct1, dct2, dct3 );
+
+ dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0,
+ ( v16u8 ) dct0, ( v16u8 ) dct0_mask );
+ dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1,
+ ( v16u8 ) dct1, ( v16u8 ) dct1_mask );
+ dct2 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h2,
+ ( v16u8 ) dct2, ( v16u8 ) dct2_mask );
+ dct3 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h3,
+ ( v16u8 ) dct3, ( v16u8 ) dct3_mask );
+
+ non_zero = HADD_SW_S32( ( v4u32 )( dct_h0 + dct_h1 + dct_h2 + dct_h3 ) );
+ ST_SH4( dct0, dct1, dct2, dct3, p_dct, 8 );
+ LD_SH4( p_dct + 32, 8, dct0, dct1, dct2, dct3 );
+
+ dct0_mask = __msa_clei_s_h( dct0, 0 );
+ dct1_mask = __msa_clei_s_h( dct1, 0 );
+ dct2_mask = __msa_clei_s_h( dct2, 0 );
+ dct3_mask = __msa_clei_s_h( dct3, 0 );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+ UNPCK_SH_SW( dct2, dct_signed_w4, dct_signed_w5 );
+ UNPCK_SH_SW( dct3, dct_signed_w6, dct_signed_w7 );
+ LD_SH4( p_bias + 32, 8, bias_h0, bias_h1, bias_h2, bias_h3 );
+ ILVR_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
+ bias0, bias2, bias4, bias6 );
+ ILVL_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
+ bias1, bias3, bias5, bias7 );
+ LD_SH4( p_mf + 32, 8, mf_h0, mf_h1, mf_h2, mf_h3 );
+ ILVR_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
+ mf_vec0, mf_vec2, mf_vec4, mf_vec6 );
+ ILVL_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
+ mf_vec1, mf_vec3, mf_vec5, mf_vec7 );
+
+ dct_w0 = __msa_add_a_w( dct_signed_w0, bias0 );
+ dct_w1 = __msa_add_a_w( dct_signed_w1, bias1 );
+ dct_w2 = __msa_add_a_w( dct_signed_w2, bias2 );
+ dct_w3 = __msa_add_a_w( dct_signed_w3, bias3 );
+ dct_w4 = __msa_add_a_w( dct_signed_w4, bias4 );
+ dct_w5 = __msa_add_a_w( dct_signed_w5, bias5 );
+ dct_w6 = __msa_add_a_w( dct_signed_w6, bias6 );
+ dct_w7 = __msa_add_a_w( dct_signed_w7, bias7 );
+
+ dct_w0 *= mf_vec0;
+ dct_w1 *= mf_vec1;
+ dct_w2 *= mf_vec2;
+ dct_w3 *= mf_vec3;
+ dct_w4 *= mf_vec4;
+ dct_w5 *= mf_vec5;
+ dct_w6 *= mf_vec6;
+ dct_w7 *= mf_vec7;
+
+ SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
+ SRA_4V( dct_w4, dct_w5, dct_w6, dct_w7, 16 );
+ PCKEV_H2_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_h0, dct_h1 );
+ PCKEV_H2_SH( dct_w5, dct_w4, dct_w7, dct_w6, dct_h2, dct_h3 );
+ SUB4( zero, dct_h0, zero, dct_h1, zero, dct_h2, zero, dct_h3,
+ dct0, dct1, dct2, dct3 );
+
+ dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0,
+ ( v16u8 ) dct0, ( v16u8 ) dct0_mask );
+ dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1,
+ ( v16u8 ) dct1, ( v16u8 ) dct1_mask );
+ dct2 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h2,
+ ( v16u8 ) dct2, ( v16u8 ) dct2_mask );
+ dct3 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h3,
+ ( v16u8 ) dct3, ( v16u8 ) dct3_mask );
+
+ non_zero += HADD_SW_S32( ( v4u32 ) ( dct_h0 + dct_h1 + dct_h2 + dct_h3 ) );
+ ST_SH4( dct0, dct1, dct2, dct3, p_dct + 32, 8 );
+
+ return !!non_zero;
+}
+
+static int32_t avc_quant_4x4_dc_msa( int16_t *p_dct, int32_t i_mf,
+ int32_t i_bias )
+{
+ int32_t non_zero = 0;
+ v8i16 dct0, dct1, dct0_mask, dct1_mask;
+ v8i16 zero = { 0 };
+ v8i16 dct_h0, dct_h1;
+ v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
+ v4i32 dct_w0, dct_w1, dct_w2, dct_w3;
+ v4i32 mf_vec, bias_vec;
+
+ LD_SH2( p_dct, 8, dct0, dct1 );
+
+ dct0_mask = __msa_clei_s_h( dct0, 0 );
+ dct1_mask = __msa_clei_s_h( dct1, 0 );
+
+ UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
+ UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
+
+ bias_vec = __msa_fill_w( i_bias );
+ mf_vec = __msa_fill_w( i_mf );
+
+ dct_w0 = __msa_add_a_w( dct_signed_w0, bias_vec );
+ dct_w1 = __msa_add_a_w( dct_signed_w1, bias_vec );
+ dct_w2 = __msa_add_a_w( dct_signed_w2, bias_vec );
+ dct_w3 = __msa_add_a_w( dct_signed_w3, bias_vec );
+
+ dct_w0 *= mf_vec;
+ dct_w1 *= mf_vec;
+ dct_w2 *= mf_vec;
+ dct_w3 *= mf_vec;
+
+ SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
+ PCKEV_H2_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_h0, dct_h1 );
+
+ dct0 = zero - dct_h0;
+ dct1 = zero - dct_h1;
+ dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0,
+ ( v16u8 ) dct0, ( v16u8 ) dct0_mask );
+ dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1,
+ ( v16u8 ) dct1, ( v16u8 ) dct1_mask );
+ non_zero = HADD_SW_S32( ( v4u32 ) ( dct_h0 + dct_h1 ) );
+
+ ST_SH2( dct0, dct1, p_dct, 8 );
+
+ return !!non_zero;
+}
+
+static int32_t avc_coeff_last64_msa( int16_t *p_src )
+{
+ uint32_t u_res;
+ v8i16 src0, src1, src2, src3, src4, src5, src6, src7;
+ v8i16 tmp_h0, tmp_h1, tmp_h2, tmp_h3, tmp_h4, tmp_h5, tmp_h6, tmp_h7;
+ v16u8 tmp0, tmp1, tmp2, tmp3;
+ v8u16 vec0, vec1, vec2, vec3;
+ v4i32 out0;
+ v16u8 mask = { 1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128 };
+
+ LD_SH8( p_src, 8, src0, src1, src2, src3, src4, src5, src6, src7 );
+
+ tmp_h0 = __msa_ceqi_h( src0, 0 );
+ tmp_h1 = __msa_ceqi_h( src1, 0 );
+ tmp_h2 = __msa_ceqi_h( src2, 0 );
+ tmp_h3 = __msa_ceqi_h( src3, 0 );
+ tmp_h4 = __msa_ceqi_h( src4, 0 );
+ tmp_h5 = __msa_ceqi_h( src5, 0 );
+ tmp_h6 = __msa_ceqi_h( src6, 0 );
+ tmp_h7 = __msa_ceqi_h( src7, 0 );
+
+ PCKEV_B4_UB( tmp_h1, tmp_h0, tmp_h3, tmp_h2, tmp_h5, tmp_h4, tmp_h7, tmp_h6,
+ tmp0, tmp1, tmp2, tmp3 );
+
+ tmp0 = tmp0 & mask;
+ tmp1 = tmp1 & mask;
+ tmp2 = tmp2 & mask;
+ tmp3 = tmp3 & mask;
+
+ HADD_UB4_UH( tmp0, tmp1, tmp2, tmp3, vec0, vec1, vec2, vec3 );
+ PCKEV_B2_UB( vec1, vec0, vec3, vec2, tmp0, tmp1 );
+ HADD_UB2_UH( tmp0, tmp1, vec0, vec1 );
+
+ tmp0 = ( v16u8 ) __msa_pckev_b( ( v16i8 ) vec1, ( v16i8 ) vec0 );
+ vec0 = __msa_hadd_u_h( tmp0, tmp0 );
+ tmp0 = ( v16u8 ) __msa_pckev_b( ( v16i8 ) vec0, ( v16i8 ) vec0 );
+ out0 = ( v4i32 ) __msa_nloc_d( ( v2i64 ) tmp0 );
+ u_res = __msa_copy_u_w( out0, 0 );
+
+ return ( 63 - u_res );
+}
+
+static int32_t avc_coeff_last16_msa( int16_t *p_src )
+{
+ uint32_t u_res;
+ v8i16 src0, src1;
+ v8u16 tmp_h0;
+ v16u8 tmp0;
+ v8i16 out0, out1;
+ v16i8 res0;
+ v16u8 mask = { 1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128 };
+
+ LD_SH2( p_src, 8, src0, src1 );
+
+ out0 = __msa_ceqi_h( src0, 0 );
+ out1 = __msa_ceqi_h( src1, 0 );
+
+ tmp0 = ( v16u8 ) __msa_pckev_b( ( v16i8 ) out1, ( v16i8 ) out0 );
+ tmp0 = tmp0 & mask;
+ tmp_h0 = __msa_hadd_u_h( tmp0, tmp0 );
+ tmp0 = ( v16u8 ) __msa_pckev_b( ( v16i8 ) tmp_h0, ( v16i8 ) tmp_h0 );
+ tmp_h0 = __msa_hadd_u_h( tmp0, tmp0 );
+ tmp0 = ( v16u8 ) __msa_pckev_b( ( v16i8 ) tmp_h0, ( v16i8 ) tmp_h0 );
+ tmp_h0 = __msa_hadd_u_h( tmp0, tmp0 );
+ res0 = __msa_pckev_b( ( v16i8 ) tmp_h0, ( v16i8 ) tmp_h0 );
+ out0 = __msa_nloc_h( ( v8i16 ) res0 );
+ u_res = __msa_copy_u_h( out0, 0 );
+
+ return ( 15 - u_res );
+}
+
+void x264_dequant_4x4_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][16],
+ int32_t i_qp )
+{
+ avc_dequant_4x4_msa( p_dct, pi_dequant_mf, i_qp );
+}
+
+void x264_dequant_8x8_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][64],
+ int32_t i_qp )
+{
+ avc_dequant_8x8_msa( p_dct, pi_dequant_mf, i_qp );
+}
+
+void x264_dequant_4x4_dc_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][16],
+ int32_t i_qp )
+{
+ avc_dequant_4x4_dc_msa( p_dct, pi_dequant_mf, i_qp );
+}
+
+int32_t x264_quant_4x4_msa( int16_t *p_dct, uint16_t *p_mf, uint16_t *p_bias )
+{
+ return avc_quant_4x4_msa( p_dct, p_mf, p_bias );
+}
+
+int32_t x264_quant_4x4x4_msa( int16_t p_dct[4][16],
+ uint16_t pu_mf[16], uint16_t pu_bias[16] )
+{
+ int32_t i_non_zero, i_non_zero_acc = 0;
+
+ for( int32_t j = 0; j < 4; j++ )
+ {
+ i_non_zero = x264_quant_4x4_msa( p_dct[j], pu_mf, pu_bias );
+
+ i_non_zero_acc |= ( !!i_non_zero ) << j;
+ }
+
+ return i_non_zero_acc;
+}
+
+int32_t x264_quant_8x8_msa( int16_t *p_dct, uint16_t *p_mf, uint16_t *p_bias )
+{
+ return avc_quant_8x8_msa( p_dct, p_mf, p_bias );
+}
+
+int32_t x264_quant_4x4_dc_msa( int16_t *p_dct, int32_t i_mf, int32_t i_bias )
+{
+ return avc_quant_4x4_dc_msa( p_dct, i_mf, i_bias );
+}
+
+int32_t x264_coeff_last64_msa( int16_t *p_src )
+{
+ return avc_coeff_last64_msa( p_src );
+}
+
+int32_t x264_coeff_last16_msa( int16_t *p_src )
+{
+ return avc_coeff_last16_msa( p_src );
+}
+#endif
diff --git a/common/mips/quant.h b/common/mips/quant.h
new file mode 100644
index 0000000..90866c2
--- /dev/null
+++ b/common/mips/quant.h
@@ -0,0 +1,43 @@
+/*****************************************************************************
+ * quant.h: msa quantization and level-run
+ *****************************************************************************
+ * Copyright (C) 2015 x264 project
+ *
+ * Authors: Rishikesh More <rishikesh.more at imgtec.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
+ *
+ * This program is also available under a commercial proprietary license.
+ * For more information, contact us at licensing at x264.com.
+ *****************************************************************************/
+
+#ifndef X264_MIPS_QUANT_H
+#define X264_MIPS_QUANT_H
+
+void x264_dequant_4x4_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][16],
+ int32_t i_qp );
+void x264_dequant_8x8_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][64],
+ int32_t i_qp );
+void x264_dequant_4x4_dc_msa( int16_t *p_dct, int32_t pi_dequant_mf[6][16],
+ int32_t i_qp );
+int32_t x264_quant_4x4_msa( int16_t *p_dct, uint16_t *p_mf, uint16_t *p_bias );
+int32_t x264_quant_4x4x4_msa( int16_t p_dct[4][16],
+ uint16_t pu_mf[16], uint16_t pu_bias[16] );
+int32_t x264_quant_8x8_msa( int16_t *p_dct, uint16_t *p_mf, uint16_t *p_bias );
+int32_t x264_quant_4x4_dc_msa( int16_t *p_dct, int32_t i_mf, int32_t i_bias );
+int32_t x264_coeff_last64_msa( int16_t *p_src );
+int32_t x264_coeff_last16_msa( int16_t *p_src );
+
+#endif
diff --git a/common/quant.c b/common/quant.c
index 318b182..bc9e8d7 100644
--- a/common/quant.c
+++ b/common/quant.c
@@ -40,6 +40,9 @@
#if ARCH_AARCH64
# include "aarch64/quant.h"
#endif
+#if ARCH_MIPS
+# include "mips/quant.h"
+#endif
#define QUANT_ONE( coef, mf, f ) \
{ \
@@ -767,6 +770,21 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
pf->denoise_dct = x264_denoise_dct_neon;
}
#endif
+
+#if HAVE_MSA
+ if( cpu&X264_CPU_MSA )
+ {
+ pf->quant_4x4 = x264_quant_4x4_msa;
+ pf->quant_4x4_dc = x264_quant_4x4_dc_msa;
+ pf->quant_4x4x4 = x264_quant_4x4x4_msa;
+ pf->quant_8x8 = x264_quant_8x8_msa;
+ pf->dequant_4x4 = x264_dequant_4x4_msa;
+ pf->dequant_4x4_dc = x264_dequant_4x4_dc_msa;
+ pf->dequant_8x8 = x264_dequant_8x8_msa;
+ pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_msa;
+ pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_msa;
+ }
+#endif
#endif // HIGH_BIT_DEPTH
pf->coeff_last[DCT_LUMA_DC] = pf->coeff_last[DCT_CHROMAU_DC] = pf->coeff_last[DCT_CHROMAV_DC] =
pf->coeff_last[DCT_CHROMAU_4x4] = pf->coeff_last[DCT_CHROMAV_4x4] = pf->coeff_last[DCT_LUMA_4x4];
--
2.3.7
More information about the x264-devel
mailing list