aarch64: NEON intrinsics dct_unquantize_h263.
authorJames Yu <james.yu@linaro.org>
Sat, 17 May 2014 04:21:10 +0000 (12:21 +0800)
committerJanne Grunau <janne-libav@jannau.net>
Mon, 21 Jul 2014 21:18:29 +0000 (23:18 +0200)
Intrinsics only used on aarch64 since the existing ARMv7 NEON asm
is slightly faster (Cortex-A9, gcc-4.8, micro-benchmarks and full
decoding time).

Signed-off-by: James Yu <james.yu@linaro.org>
Signed-off-by: Janne Grunau <janne-libav@jannau.net>
libavcodec/mpegvideo.c
libavcodec/mpegvideo.h
libavcodec/neon/Makefile [new file with mode: 0644]
libavcodec/neon/mpegvideo.c [new file with mode: 0644]

index e1daa2a..da42541 100644 (file)
@@ -393,6 +393,9 @@ av_cold int ff_dct_common_init(MpegEncContext *s)
         s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_bitexact;
     s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_c;
 
+    if (HAVE_INTRINSICS_NEON)
+        ff_MPV_common_init_neon(s);
+
     if (ARCH_ARM)
         ff_MPV_common_init_arm(s);
     if (ARCH_PPC)
index 3346509..3bd87df 100644 (file)
@@ -720,6 +720,7 @@ int ff_MPV_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
 void ff_MPV_encode_init_x86(MpegEncContext *s);
 void ff_MPV_common_init_x86(MpegEncContext *s);
 void ff_MPV_common_init_arm(MpegEncContext *s);
+void ff_MPV_common_init_neon(MpegEncContext *s);
 void ff_MPV_common_init_ppc(MpegEncContext *s);
 void ff_clean_intra_table_entries(MpegEncContext *s);
 void ff_mpeg_draw_horiz_band(MpegEncContext *s, int y, int h);
diff --git a/libavcodec/neon/Makefile b/libavcodec/neon/Makefile
new file mode 100644 (file)
index 0000000..607f116
--- /dev/null
@@ -0,0 +1 @@
+OBJS-$(CONFIG_MPEGVIDEO)          += neon/mpegvideo.o
diff --git a/libavcodec/neon/mpegvideo.c b/libavcodec/neon/mpegvideo.c
new file mode 100644 (file)
index 0000000..cb9bd66
--- /dev/null
@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) 2010 Mans Rullgard
+ * Copyright (c) 2014 James Yu <james.yu@linaro.org>
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <arm_neon.h>
+
+#include "config.h"
+
+#include "libavutil/cpu.h"
+#if   ARCH_AARCH64
+#   include "libavutil/aarch64/cpu.h"
+#elif ARCH_ARM
+#   include "libavutil/arm/cpu.h"
+#endif
+
+#include "libavcodec/mpegvideo.h"
+
+static void inline ff_dct_unquantize_h263_neon(int qscale, int qadd, int nCoeffs,
+                                               int16_t *block)
+{
+    int16x8_t q0s16, q2s16, q3s16, q8s16, q10s16, q11s16, q13s16;
+    int16x8_t q14s16, q15s16, qzs16;
+    int16x4_t d0s16, d2s16, d3s16, dzs16;
+    uint16x8_t q1u16, q9u16;
+    uint16x4_t d1u16;
+
+    dzs16 = vdup_n_s16(0);
+    qzs16 = vdupq_n_s16(0);
+
+    q15s16 = vdupq_n_s16(qscale << 1);
+    q14s16 = vdupq_n_s16(qadd);
+    q13s16 = vnegq_s16(q14s16);
+
+    if (nCoeffs > 4) {
+        for (; nCoeffs > 8; nCoeffs -= 16, block += 16) {
+            q0s16 = vld1q_s16(block);
+            q3s16 = vreinterpretq_s16_u16(vcltq_s16(q0s16, qzs16));
+            q8s16 = vld1q_s16(block + 8);
+            q1u16 = vceqq_s16(q0s16, qzs16);
+            q2s16 = vmulq_s16(q0s16, q15s16);
+            q11s16 = vreinterpretq_s16_u16(vcltq_s16(q8s16, qzs16));
+            q10s16 = vmulq_s16(q8s16, q15s16);
+            q3s16 = vbslq_s16(vreinterpretq_u16_s16(q3s16), q13s16, q14s16);
+            q11s16 = vbslq_s16(vreinterpretq_u16_s16(q11s16), q13s16, q14s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            q9u16 = vceqq_s16(q8s16, qzs16);
+            q10s16 = vaddq_s16(q10s16, q11s16);
+            q0s16 = vbslq_s16(q1u16, q0s16, q2s16);
+            q8s16 = vbslq_s16(q9u16, q8s16, q10s16);
+            vst1q_s16(block, q0s16);
+            vst1q_s16(block + 8, q8s16);
+        }
+    }
+    if (nCoeffs <= 0)
+        return;
+
+    d0s16 = vld1_s16(block);
+    d3s16 = vreinterpret_s16_u16(vclt_s16(d0s16, dzs16));
+    d1u16 = vceq_s16(d0s16, dzs16);
+    d2s16 = vmul_s16(d0s16, vget_high_s16(q15s16));
+    d3s16 = vbsl_s16(vreinterpret_u16_s16(d3s16),
+                     vget_high_s16(q13s16), vget_high_s16(q14s16));
+    d2s16 = vadd_s16(d2s16, d3s16);
+    d0s16 = vbsl_s16(d1u16, d0s16, d2s16);
+    vst1_s16(block, d0s16);
+}
+
+static void dct_unquantize_h263_inter_neon(MpegEncContext *s, int16_t *block,
+                                           int n, int qscale)
+{
+    int nCoeffs = s->inter_scantable.raster_end[s->block_last_index[n]];
+    int qadd    = (qscale - 1) | 1;
+
+    ff_dct_unquantize_h263_neon(qscale, qadd, nCoeffs + 1, block);
+}
+
+static void dct_unquantize_h263_intra_neon(MpegEncContext *s, int16_t *block,
+                                           int n, int qscale)
+{
+    int qadd;
+    int nCoeffs, blk0;
+
+    if (!s->h263_aic) {
+        if (n < 4)
+            block[0] *= s->y_dc_scale;
+        else
+            block[0] *= s->c_dc_scale;
+        qadd = (qscale - 1) | 1;
+    } else {
+        qadd = 0;
+    }
+
+    if (s->ac_pred) {
+        nCoeffs = 63;
+    } else {
+        nCoeffs = s->inter_scantable.raster_end[s->block_last_index[n]];
+        if (nCoeffs <= 0)
+            return;
+    }
+
+    blk0 = block[0];
+
+    ff_dct_unquantize_h263_neon(qscale, qadd, nCoeffs + 1, block);
+
+    block[0] = blk0;
+}
+
+
+av_cold void ff_MPV_common_init_neon(MpegEncContext *s)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+        s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_neon;
+        s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_neon;
+    }
+}