summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rwxr-xr-xconfigure1
-rw-r--r--test/android/Android.mk8
-rw-r--r--test/borders_test.cc1
-rw-r--r--test/cpu_speed_test.cc1
-rw-r--r--third_party/libwebm/Android.mk11
-rw-r--r--vp8/common/arm/neon/iwalsh_neon.asm87
-rw-r--r--vp8/common/arm/neon/iwalsh_neon.c102
-rw-r--r--vp8/common/arm/neon/loopfilter_neon.asm409
-rw-r--r--vp8/common/arm/neon/loopfilter_neon.c538
-rw-r--r--vp8/encoder/arm/neon/denoising_neon.c59
-rw-r--r--vp8/encoder/x86/quantize_ssse3.asm138
-rw-r--r--vp8/encoder/x86/quantize_ssse3.c110
-rw-r--r--vp8/vp8_common.mk4
-rw-r--r--vp8/vp8cx.mk2
-rw-r--r--vp9/encoder/vp9_encoder.h1
-rw-r--r--vp9/encoder/vp9_rdopt.c3
-rw-r--r--vp9/encoder/vp9_temporal_filter.c126
-rw-r--r--vp9/vp9_cx_iface.c6
-rw-r--r--vpx/vp8cx.h6
-rw-r--r--vpxenc.c4
20 files changed, 911 insertions, 706 deletions
diff --git a/configure b/configure
index 29683cd62..690ac48fa 100755
--- a/configure
+++ b/configure
@@ -709,6 +709,7 @@ process_toolchain() {
soft_enable webm_io
;;
*-android-*)
+ soft_enable webm_io
# GTestLog must be modified to use Android logging utilities.
;;
*-darwin-*)
diff --git a/test/android/Android.mk b/test/android/Android.mk
index 13af601e2..4e750b24b 100644
--- a/test/android/Android.mk
+++ b/test/android/Android.mk
@@ -14,8 +14,14 @@ CUR_WD := $(call my-dir)
BINDINGS_DIR := $(CUR_WD)/../../..
LOCAL_PATH := $(CUR_WD)/../../..
+#libwebm
+include $(CLEAR_VARS)
+include $(BINDINGS_DIR)/libvpx/third_party/libwebm/Android.mk
+LOCAL_PATH := $(CUR_WD)/../../..
+
#libvpx
include $(CLEAR_VARS)
+LOCAL_STATIC_LIBRARIES := libwebm
include $(BINDINGS_DIR)/libvpx/build/make/Android.mk
LOCAL_PATH := $(CUR_WD)/../..
@@ -33,7 +39,7 @@ include $(BUILD_STATIC_LIBRARY)
include $(CLEAR_VARS)
LOCAL_ARM_MODE := arm
LOCAL_MODULE := libvpx_test
-LOCAL_STATIC_LIBRARIES := gtest
+LOCAL_STATIC_LIBRARIES := gtest libwebm
LOCAL_SHARED_LIBRARIES := vpx
include $(LOCAL_PATH)/test/test.mk
LOCAL_C_INCLUDES := $(BINDINGS_DIR)
diff --git a/test/borders_test.cc b/test/borders_test.cc
index a2f5a1bdd..b30be4580 100644
--- a/test/borders_test.cc
+++ b/test/borders_test.cc
@@ -35,6 +35,7 @@ class BordersTest : public ::libvpx_test::EncoderTest,
encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+ encoder->Control(VP8E_SET_ARNR_TYPE, 3);
}
}
diff --git a/test/cpu_speed_test.cc b/test/cpu_speed_test.cc
index ca201bbc7..be651b4fe 100644
--- a/test/cpu_speed_test.cc
+++ b/test/cpu_speed_test.cc
@@ -37,6 +37,7 @@ class CpuSpeedTest : public ::libvpx_test::EncoderTest,
encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+ encoder->Control(VP8E_SET_ARNR_TYPE, 3);
}
}
diff --git a/third_party/libwebm/Android.mk b/third_party/libwebm/Android.mk
new file mode 100644
index 000000000..13868b660
--- /dev/null
+++ b/third_party/libwebm/Android.mk
@@ -0,0 +1,11 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+LOCAL_CPP_EXTENSION := .cpp
+LOCAL_SRC_FILES := mkvmuxer.cpp \
+ mkvmuxerutil.cpp \
+ mkvparser.cpp \
+ mkvreader.cpp \
+ mkvwriter.cpp
+LOCAL_MODULE := libwebm
+include $(BUILD_STATIC_LIBRARY)
diff --git a/vp8/common/arm/neon/iwalsh_neon.asm b/vp8/common/arm/neon/iwalsh_neon.asm
deleted file mode 100644
index e8ea2a619..000000000
--- a/vp8/common/arm/neon/iwalsh_neon.asm
+++ /dev/null
@@ -1,87 +0,0 @@
-;
-; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
- EXPORT |vp8_short_inv_walsh4x4_neon|
-
- ARM
- REQUIRE8
- PRESERVE8
-
- AREA |.text|, CODE, READONLY ; name this block of code
-
-;short vp8_short_inv_walsh4x4_neon(short *input, short *mb_dqcoeff)
-|vp8_short_inv_walsh4x4_neon| PROC
-
- ; read in all four lines of values: d0->d3
- vld1.i16 {q0-q1}, [r0@128]
-
- ; first for loop
- vadd.s16 d4, d0, d3 ;a = [0] + [12]
- vadd.s16 d6, d1, d2 ;b = [4] + [8]
- vsub.s16 d5, d0, d3 ;d = [0] - [12]
- vsub.s16 d7, d1, d2 ;c = [4] - [8]
-
- vadd.s16 q0, q2, q3 ; a+b d+c
- vsub.s16 q1, q2, q3 ; a-b d-c
-
- vtrn.32 d0, d2 ;d0: 0 1 8 9
- ;d2: 2 3 10 11
- vtrn.32 d1, d3 ;d1: 4 5 12 13
- ;d3: 6 7 14 15
-
- vtrn.16 d0, d1 ;d0: 0 4 8 12
- ;d1: 1 5 9 13
- vtrn.16 d2, d3 ;d2: 2 6 10 14
- ;d3: 3 7 11 15
-
- ; second for loop
-
- vadd.s16 d4, d0, d3 ;a = [0] + [3]
- vadd.s16 d6, d1, d2 ;b = [1] + [2]
- vsub.s16 d5, d0, d3 ;d = [0] - [3]
- vsub.s16 d7, d1, d2 ;c = [1] - [2]
-
- vmov.i16 q8, #3
-
- vadd.s16 q0, q2, q3 ; a+b d+c
- vsub.s16 q1, q2, q3 ; a-b d-c
-
- vadd.i16 q0, q0, q8 ;e/f += 3
- vadd.i16 q1, q1, q8 ;g/h += 3
-
- vshr.s16 q0, q0, #3 ;e/f >> 3
- vshr.s16 q1, q1, #3 ;g/h >> 3
-
- mov r2, #64
- add r3, r1, #32
-
- vst1.i16 d0[0], [r1],r2
- vst1.i16 d1[0], [r3],r2
- vst1.i16 d2[0], [r1],r2
- vst1.i16 d3[0], [r3],r2
-
- vst1.i16 d0[1], [r1],r2
- vst1.i16 d1[1], [r3],r2
- vst1.i16 d2[1], [r1],r2
- vst1.i16 d3[1], [r3],r2
-
- vst1.i16 d0[2], [r1],r2
- vst1.i16 d1[2], [r3],r2
- vst1.i16 d2[2], [r1],r2
- vst1.i16 d3[2], [r3],r2
-
- vst1.i16 d0[3], [r1],r2
- vst1.i16 d1[3], [r3],r2
- vst1.i16 d2[3], [r1]
- vst1.i16 d3[3], [r3]
-
- bx lr
- ENDP ; |vp8_short_inv_walsh4x4_neon|
-
- END
diff --git a/vp8/common/arm/neon/iwalsh_neon.c b/vp8/common/arm/neon/iwalsh_neon.c
new file mode 100644
index 000000000..6ea9dd712
--- /dev/null
+++ b/vp8/common/arm/neon/iwalsh_neon.c
@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_short_inv_walsh4x4_neon(
+ int16_t *input,
+ int16_t *mb_dqcoeff) {
+ int16x8_t q0s16, q1s16, q2s16, q3s16;
+ int16x4_t d4s16, d5s16, d6s16, d7s16;
+ int16x4x2_t v2tmp0, v2tmp1;
+ int32x2x2_t v2tmp2, v2tmp3;
+ int16x8_t qAdd3;
+
+ q0s16 = vld1q_s16(input);
+ q1s16 = vld1q_s16(input + 8);
+
+ // 1st for loop
+ d4s16 = vadd_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+ d6s16 = vadd_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+ d5s16 = vsub_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+ d7s16 = vsub_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+
+ q2s16 = vcombine_s16(d4s16, d5s16);
+ q3s16 = vcombine_s16(d6s16, d7s16);
+
+ q0s16 = vaddq_s16(q2s16, q3s16);
+ q1s16 = vsubq_s16(q2s16, q3s16);
+
+ v2tmp2 = vtrn_s32(vreinterpret_s32_s16(vget_low_s16(q0s16)),
+ vreinterpret_s32_s16(vget_low_s16(q1s16)));
+ v2tmp3 = vtrn_s32(vreinterpret_s32_s16(vget_high_s16(q0s16)),
+ vreinterpret_s32_s16(vget_high_s16(q1s16)));
+ v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),
+ vreinterpret_s16_s32(v2tmp3.val[0]));
+ v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),
+ vreinterpret_s16_s32(v2tmp3.val[1]));
+
+ // 2nd for loop
+ d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
+ d6s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
+ d5s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
+ d7s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
+ q2s16 = vcombine_s16(d4s16, d5s16);
+ q3s16 = vcombine_s16(d6s16, d7s16);
+
+ qAdd3 = vdupq_n_s16(3);
+
+ q0s16 = vaddq_s16(q2s16, q3s16);
+ q1s16 = vsubq_s16(q2s16, q3s16);
+
+ q0s16 = vaddq_s16(q0s16, qAdd3);
+ q1s16 = vaddq_s16(q1s16, qAdd3);
+
+ q0s16 = vshrq_n_s16(q0s16, 3);
+ q1s16 = vshrq_n_s16(q1s16, 3);
+
+ // store
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16), 0);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 0);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16), 0);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 0);
+ mb_dqcoeff += 16;
+
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16), 1);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 1);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16), 1);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 1);
+ mb_dqcoeff += 16;
+
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16), 2);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 2);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16), 2);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 2);
+ mb_dqcoeff += 16;
+
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16), 3);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 3);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16), 3);
+ mb_dqcoeff += 16;
+ vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 3);
+ mb_dqcoeff += 16;
+ return;
+}
diff --git a/vp8/common/arm/neon/loopfilter_neon.asm b/vp8/common/arm/neon/loopfilter_neon.asm
deleted file mode 100644
index c4f09c775..000000000
--- a/vp8/common/arm/neon/loopfilter_neon.asm
+++ /dev/null
@@ -1,409 +0,0 @@
-;
-; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
-
- EXPORT |vp8_loop_filter_horizontal_edge_y_neon|
- EXPORT |vp8_loop_filter_horizontal_edge_uv_neon|
- EXPORT |vp8_loop_filter_vertical_edge_y_neon|
- EXPORT |vp8_loop_filter_vertical_edge_uv_neon|
- ARM
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; r0 unsigned char *src
-; r1 int pitch
-; r2 unsigned char blimit
-; r3 unsigned char limit
-; sp unsigned char thresh,
-|vp8_loop_filter_horizontal_edge_y_neon| PROC
- push {lr}
- vpush {d8-d15}
-
- vdup.u8 q0, r2 ; duplicate blimit
- vdup.u8 q1, r3 ; duplicate limit
- sub r2, r0, r1, lsl #2 ; move src pointer down by 4 lines
- ldr r3, [sp, #68] ; load thresh
- add r12, r2, r1
- add r1, r1, r1
-
- vdup.u8 q2, r3 ; duplicate thresh
-
- vld1.u8 {q3}, [r2@128], r1 ; p3
- vld1.u8 {q4}, [r12@128], r1 ; p2
- vld1.u8 {q5}, [r2@128], r1 ; p1
- vld1.u8 {q6}, [r12@128], r1 ; p0
- vld1.u8 {q7}, [r2@128], r1 ; q0
- vld1.u8 {q8}, [r12@128], r1 ; q1
- vld1.u8 {q9}, [r2@128] ; q2
- vld1.u8 {q10}, [r12@128] ; q3
-
- sub r2, r2, r1, lsl #1
- sub r12, r12, r1, lsl #1
-
- bl vp8_loop_filter_neon
-
- vst1.u8 {q5}, [r2@128], r1 ; store op1
- vst1.u8 {q6}, [r12@128], r1 ; store op0
- vst1.u8 {q7}, [r2@128], r1 ; store oq0
- vst1.u8 {q8}, [r12@128], r1 ; store oq1
-
- vpop {d8-d15}
- pop {pc}
- ENDP ; |vp8_loop_filter_horizontal_edge_y_neon|
-
-
-; r0 unsigned char *u,
-; r1 int pitch,
-; r2 unsigned char blimit
-; r3 unsigned char limit
-; sp unsigned char thresh,
-; sp+4 unsigned char *v
-|vp8_loop_filter_horizontal_edge_uv_neon| PROC
- push {lr}
- vpush {d8-d15}
-
- vdup.u8 q0, r2 ; duplicate blimit
- vdup.u8 q1, r3 ; duplicate limit
- ldr r12, [sp, #68] ; load thresh
- ldr r2, [sp, #72] ; load v ptr
- vdup.u8 q2, r12 ; duplicate thresh
-
- sub r3, r0, r1, lsl #2 ; move u pointer down by 4 lines
- sub r12, r2, r1, lsl #2 ; move v pointer down by 4 lines
-
- vld1.u8 {d6}, [r3@64], r1 ; p3
- vld1.u8 {d7}, [r12@64], r1 ; p3
- vld1.u8 {d8}, [r3@64], r1 ; p2
- vld1.u8 {d9}, [r12@64], r1 ; p2
- vld1.u8 {d10}, [r3@64], r1 ; p1
- vld1.u8 {d11}, [r12@64], r1 ; p1
- vld1.u8 {d12}, [r3@64], r1 ; p0
- vld1.u8 {d13}, [r12@64], r1 ; p0
- vld1.u8 {d14}, [r3@64], r1 ; q0
- vld1.u8 {d15}, [r12@64], r1 ; q0
- vld1.u8 {d16}, [r3@64], r1 ; q1
- vld1.u8 {d17}, [r12@64], r1 ; q1
- vld1.u8 {d18}, [r3@64], r1 ; q2
- vld1.u8 {d19}, [r12@64], r1 ; q2
- vld1.u8 {d20}, [r3@64] ; q3
- vld1.u8 {d21}, [r12@64] ; q3
-
- bl vp8_loop_filter_neon
-
- sub r0, r0, r1, lsl #1
- sub r2, r2, r1, lsl #1
-
- vst1.u8 {d10}, [r0@64], r1 ; store u op1
- vst1.u8 {d11}, [r2@64], r1 ; store v op1
- vst1.u8 {d12}, [r0@64], r1 ; store u op0
- vst1.u8 {d13}, [r2@64], r1 ; store v op0
- vst1.u8 {d14}, [r0@64], r1 ; store u oq0
- vst1.u8 {d15}, [r2@64], r1 ; store v oq0
- vst1.u8 {d16}, [r0@64] ; store u oq1
- vst1.u8 {d17}, [r2@64] ; store v oq1
-
- vpop {d8-d15}
- pop {pc}
- ENDP ; |vp8_loop_filter_horizontal_edge_uv_neon|
-
-; void vp8_loop_filter_vertical_edge_y_neon(unsigned char *src, int pitch,
-; const signed char *flimit,
-; const signed char *limit,
-; const signed char *thresh,
-; int count)
-; r0 unsigned char *src
-; r1 int pitch
-; r2 unsigned char blimit
-; r3 unsigned char limit
-; sp unsigned char thresh,
-
-|vp8_loop_filter_vertical_edge_y_neon| PROC
- push {lr}
- vpush {d8-d15}
-
- vdup.u8 q0, r2 ; duplicate blimit
- vdup.u8 q1, r3 ; duplicate limit
- sub r2, r0, #4 ; src ptr down by 4 columns
- add r1, r1, r1
- ldr r3, [sp, #68] ; load thresh
- add r12, r2, r1, asr #1
-
- vld1.u8 {d6}, [r2], r1
- vld1.u8 {d8}, [r12], r1
- vld1.u8 {d10}, [r2], r1
- vld1.u8 {d12}, [r12], r1
- vld1.u8 {d14}, [r2], r1
- vld1.u8 {d16}, [r12], r1
- vld1.u8 {d18}, [r2], r1
- vld1.u8 {d20}, [r12], r1
-
- vld1.u8 {d7}, [r2], r1 ; load second 8-line src data
- vld1.u8 {d9}, [r12], r1
- vld1.u8 {d11}, [r2], r1
- vld1.u8 {d13}, [r12], r1
- vld1.u8 {d15}, [r2], r1
- vld1.u8 {d17}, [r12], r1
- vld1.u8 {d19}, [r2]
- vld1.u8 {d21}, [r12]
-
- ;transpose to 8x16 matrix
- vtrn.32 q3, q7
- vtrn.32 q4, q8
- vtrn.32 q5, q9
- vtrn.32 q6, q10
-
- vdup.u8 q2, r3 ; duplicate thresh
-
- vtrn.16 q3, q5
- vtrn.16 q4, q6
- vtrn.16 q7, q9
- vtrn.16 q8, q10
-
- vtrn.8 q3, q4
- vtrn.8 q5, q6
- vtrn.8 q7, q8
- vtrn.8 q9, q10
-
- bl vp8_loop_filter_neon
-
- vswp d12, d11
- vswp d16, d13
-
- sub r0, r0, #2 ; dst ptr
-
- vswp d14, d12
- vswp d16, d15
-
- add r12, r0, r1, asr #1
-
- ;store op1, op0, oq0, oq1
- vst4.8 {d10[0], d11[0], d12[0], d13[0]}, [r0], r1
- vst4.8 {d10[1], d11[1], d12[1], d13[1]}, [r12], r1
- vst4.8 {d10[2], d11[2], d12[2], d13[2]}, [r0], r1
- vst4.8 {d10[3], d11[3], d12[3], d13[3]}, [r12], r1
- vst4.8 {d10[4], d11[4], d12[4], d13[4]}, [r0], r1
- vst4.8 {d10[5], d11[5], d12[5], d13[5]}, [r12], r1
- vst4.8 {d10[6], d11[6], d12[6], d13[6]}, [r0], r1
- vst4.8 {d10[7], d11[7], d12[7], d13[7]}, [r12], r1
-
- vst4.8 {d14[0], d15[0], d16[0], d17[0]}, [r0], r1
- vst4.8 {d14[1], d15[1], d16[1], d17[1]}, [r12], r1
- vst4.8 {d14[2], d15[2], d16[2], d17[2]}, [r0], r1
- vst4.8 {d14[3], d15[3], d16[3], d17[3]}, [r12], r1
- vst4.8 {d14[4], d15[4], d16[4], d17[4]}, [r0], r1
- vst4.8 {d14[5], d15[5], d16[5], d17[5]}, [r12], r1
- vst4.8 {d14[6], d15[6], d16[6], d17[6]}, [r0]
- vst4.8 {d14[7], d15[7], d16[7], d17[7]}, [r12]
-
- vpop {d8-d15}
- pop {pc}
- ENDP ; |vp8_loop_filter_vertical_edge_y_neon|
-
-; void vp8_loop_filter_vertical_edge_uv_neon(unsigned char *u, int pitch
-; const signed char *flimit,
-; const signed char *limit,
-; const signed char *thresh,
-; unsigned char *v)
-; r0 unsigned char *u,
-; r1 int pitch,
-; r2 unsigned char blimit
-; r3 unsigned char limit
-; sp unsigned char thresh,
-; sp+4 unsigned char *v
-|vp8_loop_filter_vertical_edge_uv_neon| PROC
- push {lr}
- vpush {d8-d15}
-
- vdup.u8 q0, r2 ; duplicate blimit
- sub r12, r0, #4 ; move u pointer down by 4 columns
- ldr r2, [sp, #72] ; load v ptr
- vdup.u8 q1, r3 ; duplicate limit
- sub r3, r2, #4 ; move v pointer down by 4 columns
-
- vld1.u8 {d6}, [r12], r1 ;load u data
- vld1.u8 {d7}, [r3], r1 ;load v data
- vld1.u8 {d8}, [r12], r1
- vld1.u8 {d9}, [r3], r1
- vld1.u8 {d10}, [r12], r1
- vld1.u8 {d11}, [r3], r1
- vld1.u8 {d12}, [r12], r1
- vld1.u8 {d13}, [r3], r1
- vld1.u8 {d14}, [r12], r1
- vld1.u8 {d15}, [r3], r1
- vld1.u8 {d16}, [r12], r1
- vld1.u8 {d17}, [r3], r1
- vld1.u8 {d18}, [r12], r1
- vld1.u8 {d19}, [r3], r1
- vld1.u8 {d20}, [r12]
- vld1.u8 {d21}, [r3]
-
- ldr r12, [sp, #68] ; load thresh
-
- ;transpose to 8x16 matrix
- vtrn.32 q3, q7
- vtrn.32 q4, q8
- vtrn.32 q5, q9
- vtrn.32 q6, q10
-
- vdup.u8 q2, r12 ; duplicate thresh
-
- vtrn.16 q3, q5
- vtrn.16 q4, q6
- vtrn.16 q7, q9
- vtrn.16 q8, q10
-
- vtrn.8 q3, q4
- vtrn.8 q5, q6
- vtrn.8 q7, q8
- vtrn.8 q9, q10
-
- bl vp8_loop_filter_neon
-
- vswp d12, d11
- vswp d16, d13
- vswp d14, d12
- vswp d16, d15
-
- sub r0, r0, #2
- sub r2, r2, #2
-
- ;store op1, op0, oq0, oq1
- vst4.8 {d10[0], d11[0], d12[0], d13[0]}, [r0], r1
- vst4.8 {d14[0], d15[0], d16[0], d17[0]}, [r2], r1
- vst4.8 {d10[1], d11[1], d12[1], d13[1]}, [r0], r1
- vst4.8 {d14[1], d15[1], d16[1], d17[1]}, [r2], r1
- vst4.8 {d10[2], d11[2], d12[2], d13[2]}, [r0], r1
- vst4.8 {d14[2], d15[2], d16[2], d17[2]}, [r2], r1
- vst4.8 {d10[3], d11[3], d12[3], d13[3]}, [r0], r1
- vst4.8 {d14[3], d15[3], d16[3], d17[3]}, [r2], r1
- vst4.8 {d10[4], d11[4], d12[4], d13[4]}, [r0], r1
- vst4.8 {d14[4], d15[4], d16[4], d17[4]}, [r2], r1
- vst4.8 {d10[5], d11[5], d12[5], d13[5]}, [r0], r1
- vst4.8 {d14[5], d15[5], d16[5], d17[5]}, [r2], r1
- vst4.8 {d10[6], d11[6], d12[6], d13[6]}, [r0], r1
- vst4.8 {d14[6], d15[6], d16[6], d17[6]}, [r2], r1
- vst4.8 {d10[7], d11[7], d12[7], d13[7]}, [r0]
- vst4.8 {d14[7], d15[7], d16[7], d17[7]}, [r2]
-
- vpop {d8-d15}
- pop {pc}
- ENDP ; |vp8_loop_filter_vertical_edge_uv_neon|
-
-; void vp8_loop_filter_neon();
-; This is a helper function for the loopfilters. The invidual functions do the
-; necessary load, transpose (if necessary) and store.
-
-; r0-r3 PRESERVE
-; q0 flimit
-; q1 limit
-; q2 thresh
-; q3 p3
-; q4 p2
-; q5 p1
-; q6 p0
-; q7 q0
-; q8 q1
-; q9 q2
-; q10 q3
-|vp8_loop_filter_neon| PROC
-
- ; vp8_filter_mask
- vabd.u8 q11, q3, q4 ; abs(p3 - p2)
- vabd.u8 q12, q4, q5 ; abs(p2 - p1)
- vabd.u8 q13, q5, q6 ; abs(p1 - p0)
- vabd.u8 q14, q8, q7 ; abs(q1 - q0)
- vabd.u8 q3, q9, q8 ; abs(q2 - q1)
- vabd.u8 q4, q10, q9 ; abs(q3 - q2)
-
- vmax.u8 q11, q11, q12
- vmax.u8 q12, q13, q14
- vmax.u8 q3, q3, q4
- vmax.u8 q15, q11, q12
-
- vabd.u8 q9, q6, q7 ; abs(p0 - q0)
-
- ; vp8_hevmask
- vcgt.u8 q13, q13, q2 ; (abs(p1 - p0) > thresh)*-1
- vcgt.u8 q14, q14, q2 ; (abs(q1 - q0) > thresh)*-1
- vmax.u8 q15, q15, q3
-
- vmov.u8 q10, #0x80 ; 0x80
-
- vabd.u8 q2, q5, q8 ; a = abs(p1 - q1)
- vqadd.u8 q9, q9, q9 ; b = abs(p0 - q0) * 2
-
- vcge.u8 q15, q1, q15
-
- ; vp8_filter() function
- ; convert to signed
- veor q7, q7, q10 ; qs0
- vshr.u8 q2, q2, #1 ; a = a / 2
- veor q6, q6, q10 ; ps0
-
- veor q5, q5, q10 ; ps1
- vqadd.u8 q9, q9, q2 ; a = b + a
-
- veor q8, q8, q10 ; qs1
-
- vmov.u8 q10, #3 ; #3
-
- vsubl.s8 q2, d14, d12 ; ( qs0 - ps0)
- vsubl.s8 q11, d15, d13
-
- vcge.u8 q9, q0, q9 ; (a > flimit * 2 + limit) * -1
-
- vmovl.u8 q4, d20
-
- vqsub.s8 q1, q5, q8 ; vp8_filter = clamp(ps1-qs1)
- vorr q14, q13, q14 ; vp8_hevmask
-
- vmul.i16 q2, q2, q4 ; 3 * ( qs0 - ps0)
- vmul.i16 q11, q11, q4
-
- vand q1, q1, q14 ; vp8_filter &= hev
- vand q15, q15, q9 ; vp8_filter_mask
-
- vaddw.s8 q2, q2, d2
- vaddw.s8 q11, q11, d3
-
- vmov.u8 q9, #4 ; #4
-
- ; vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
- vqmovn.s16 d2, q2
- vqmovn.s16 d3, q11
- vand q1, q1, q15 ; vp8_filter &= mask
-
- vqadd.s8 q2, q1, q10 ; Filter2 = clamp(vp8_filter+3)
- vqadd.s8 q1, q1, q9 ; Filter1 = clamp(vp8_filter+4)
- vshr.s8 q2, q2, #3 ; Filter2 >>= 3
- vshr.s8 q1, q1, #3 ; Filter1 >>= 3
-
-
- vqadd.s8 q11, q6, q2 ; u = clamp(ps0 + Filter2)
- vqsub.s8 q10, q7, q1 ; u = clamp(qs0 - Filter1)
-
- ; outer tap adjustments: ++vp8_filter >> 1
- vrshr.s8 q1, q1, #1
- vbic q1, q1, q14 ; vp8_filter &= ~hev
- vmov.u8 q0, #0x80 ; 0x80
- vqadd.s8 q13, q5, q1 ; u = clamp(ps1 + vp8_filter)
- vqsub.s8 q12, q8, q1 ; u = clamp(qs1 - vp8_filter)
-
- veor q6, q11, q0 ; *op0 = u^0x80
- veor q7, q10, q0 ; *oq0 = u^0x80
- veor q5, q13, q0 ; *op1 = u^0x80
- veor q8, q12, q0 ; *oq1 = u^0x80
-
- bx lr
- ENDP ; |vp8_loop_filter_horizontal_edge_y_neon|
-
-;-----------------
-
- END
diff --git a/vp8/common/arm/neon/loopfilter_neon.c b/vp8/common/arm/neon/loopfilter_neon.c
new file mode 100644
index 000000000..47f522bc6
--- /dev/null
+++ b/vp8/common/arm/neon/loopfilter_neon.c
@@ -0,0 +1,538 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static inline void vp8_loop_filter_neon(
+ uint8x16_t qblimit, // flimit
+ uint8x16_t qlimit, // limit
+ uint8x16_t qthresh, // thresh
+ uint8x16_t q3, // p3
+ uint8x16_t q4, // p2
+ uint8x16_t q5, // p1
+ uint8x16_t q6, // p0
+ uint8x16_t q7, // q0
+ uint8x16_t q8, // q1
+ uint8x16_t q9, // q2
+ uint8x16_t q10, // q3
+ uint8x16_t *q5r, // p1
+ uint8x16_t *q6r, // p0
+ uint8x16_t *q7r, // q0
+ uint8x16_t *q8r) { // q1
+ uint8x16_t q0u8, q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+ int16x8_t q2s16, q11s16;
+ uint16x8_t q4u16;
+ int8x16_t q1s8, q2s8, q10s8, q11s8, q12s8, q13s8;
+ int8x8_t d2s8, d3s8;
+
+ q11u8 = vabdq_u8(q3, q4);
+ q12u8 = vabdq_u8(q4, q5);
+ q13u8 = vabdq_u8(q5, q6);
+ q14u8 = vabdq_u8(q8, q7);
+ q3 = vabdq_u8(q9, q8);
+ q4 = vabdq_u8(q10, q9);
+
+ q11u8 = vmaxq_u8(q11u8, q12u8);
+ q12u8 = vmaxq_u8(q13u8, q14u8);
+ q3 = vmaxq_u8(q3, q4);
+ q15u8 = vmaxq_u8(q11u8, q12u8);
+
+ q9 = vabdq_u8(q6, q7);
+
+ // vp8_hevmask
+ q13u8 = vcgtq_u8(q13u8, qthresh);
+ q14u8 = vcgtq_u8(q14u8, qthresh);
+ q15u8 = vmaxq_u8(q15u8, q3);
+
+ q2u8 = vabdq_u8(q5, q8);
+ q9 = vqaddq_u8(q9, q9);
+
+ q15u8 = vcgeq_u8(qlimit, q15u8);
+
+ // vp8_filter() function
+ // convert to signed
+ q10 = vdupq_n_u8(0x80);
+ q8 = veorq_u8(q8, q10);
+ q7 = veorq_u8(q7, q10);
+ q6 = veorq_u8(q6, q10);
+ q5 = veorq_u8(q5, q10);
+
+ q2u8 = vshrq_n_u8(q2u8, 1);
+ q9 = vqaddq_u8(q9, q2u8);
+
+ q10 = vdupq_n_u8(3);
+
+ q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+ vget_low_s8(vreinterpretq_s8_u8(q6)));
+ q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+ vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+ q9 = vcgeq_u8(qblimit, q9);
+
+ q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+ vreinterpretq_s8_u8(q8));
+
+ q14u8 = vorrq_u8(q13u8, q14u8);
+
+ q4u16 = vmovl_u8(vget_low_u8(q10));
+ q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+ q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+ q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+ q15u8 = vandq_u8(q15u8, q9);
+
+ q1s8 = vreinterpretq_s8_u8(q1u8);
+ q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+ q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+ q9 = vdupq_n_u8(4);
+ // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+ d2s8 = vqmovn_s16(q2s16);
+ d3s8 = vqmovn_s16(q11s16);
+ q1s8 = vcombine_s8(d2s8, d3s8);
+ q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+ q1s8 = vreinterpretq_s8_u8(q1u8);
+
+ q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q10));
+ q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+ q2s8 = vshrq_n_s8(q2s8, 3);
+ q1s8 = vshrq_n_s8(q1s8, 3);
+
+ q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+ q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+ q1s8 = vrshrq_n_s8(q1s8, 1);
+ q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+ q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+ q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+ q0u8 = vdupq_n_u8(0x80);
+ *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q0u8);
+ *q7r = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
+ *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
+ *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q0u8);
+ return;
+}
+
+void vp8_loop_filter_horizontal_edge_y_neon(
+ unsigned char *src,
+ int pitch,
+ unsigned char blimit,
+ unsigned char limit,
+ unsigned char thresh) {
+ uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+ uint8x16_t q5, q6, q7, q8, q9, q10;
+
+ qblimit = vdupq_n_u8(blimit);
+ qlimit = vdupq_n_u8(limit);
+ qthresh = vdupq_n_u8(thresh);
+ src -= (pitch << 2);
+
+ q3 = vld1q_u8(src);
+ src += pitch;
+ q4 = vld1q_u8(src);
+ src += pitch;
+ q5 = vld1q_u8(src);
+ src += pitch;
+ q6 = vld1q_u8(src);
+ src += pitch;
+ q7 = vld1q_u8(src);
+ src += pitch;
+ q8 = vld1q_u8(src);
+ src += pitch;
+ q9 = vld1q_u8(src);
+ src += pitch;
+ q10 = vld1q_u8(src);
+
+ vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+ q5, q6, q7, q8, q9, q10,
+ &q5, &q6, &q7, &q8);
+
+ src -= (pitch * 5);
+ vst1q_u8(src, q5);
+ src += pitch;
+ vst1q_u8(src, q6);
+ src += pitch;
+ vst1q_u8(src, q7);
+ src += pitch;
+ vst1q_u8(src, q8);
+ return;
+}
+
+void vp8_loop_filter_horizontal_edge_uv_neon(
+ unsigned char *u,
+ int pitch,
+ unsigned char blimit,
+ unsigned char limit,
+ unsigned char thresh,
+ unsigned char *v) {
+ uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+ uint8x16_t q5, q6, q7, q8, q9, q10;
+ uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+ uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+
+ qblimit = vdupq_n_u8(blimit);
+ qlimit = vdupq_n_u8(limit);
+ qthresh = vdupq_n_u8(thresh);
+
+ u -= (pitch << 2);
+ v -= (pitch << 2);
+
+ d6 = vld1_u8(u);
+ u += pitch;
+ d7 = vld1_u8(v);
+ v += pitch;
+ d8 = vld1_u8(u);
+ u += pitch;
+ d9 = vld1_u8(v);
+ v += pitch;
+ d10 = vld1_u8(u);
+ u += pitch;
+ d11 = vld1_u8(v);
+ v += pitch;
+ d12 = vld1_u8(u);
+ u += pitch;
+ d13 = vld1_u8(v);
+ v += pitch;
+ d14 = vld1_u8(u);
+ u += pitch;
+ d15 = vld1_u8(v);
+ v += pitch;
+ d16 = vld1_u8(u);
+ u += pitch;
+ d17 = vld1_u8(v);
+ v += pitch;
+ d18 = vld1_u8(u);
+ u += pitch;
+ d19 = vld1_u8(v);
+ v += pitch;
+ d20 = vld1_u8(u);
+ d21 = vld1_u8(v);
+
+ q3 = vcombine_u8(d6, d7);
+ q4 = vcombine_u8(d8, d9);
+ q5 = vcombine_u8(d10, d11);
+ q6 = vcombine_u8(d12, d13);
+ q7 = vcombine_u8(d14, d15);
+ q8 = vcombine_u8(d16, d17);
+ q9 = vcombine_u8(d18, d19);
+ q10 = vcombine_u8(d20, d21);
+
+ vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+ q5, q6, q7, q8, q9, q10,
+ &q5, &q6, &q7, &q8);
+
+ u -= (pitch * 5);
+ vst1_u8(u, vget_low_u8(q5));
+ u += pitch;
+ vst1_u8(u, vget_low_u8(q6));
+ u += pitch;
+ vst1_u8(u, vget_low_u8(q7));
+ u += pitch;
+ vst1_u8(u, vget_low_u8(q8));
+
+ v -= (pitch * 5);
+ vst1_u8(v, vget_high_u8(q5));
+ v += pitch;
+ vst1_u8(v, vget_high_u8(q6));
+ v += pitch;
+ vst1_u8(v, vget_high_u8(q7));
+ v += pitch;
+ vst1_u8(v, vget_high_u8(q8));
+ return;
+}
+
+void vp8_loop_filter_vertical_edge_y_neon(
+ unsigned char *src,
+ int pitch,
+ unsigned char blimit,
+ unsigned char limit,
+ unsigned char thresh) {
+ unsigned char *s, *d;
+ uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+ uint8x16_t q5, q6, q7, q8, q9, q10;
+ uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+ uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+ uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+ uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+ uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+ uint8x8x4_t q4ResultH, q4ResultL;
+
+ qblimit = vdupq_n_u8(blimit);
+ qlimit = vdupq_n_u8(limit);
+ qthresh = vdupq_n_u8(thresh);
+
+ s = src - 4;
+ d6 = vld1_u8(s);
+ s += pitch;
+ d8 = vld1_u8(s);
+ s += pitch;
+ d10 = vld1_u8(s);
+ s += pitch;
+ d12 = vld1_u8(s);
+ s += pitch;
+ d14 = vld1_u8(s);
+ s += pitch;
+ d16 = vld1_u8(s);
+ s += pitch;
+ d18 = vld1_u8(s);
+ s += pitch;
+ d20 = vld1_u8(s);
+ s += pitch;
+ d7 = vld1_u8(s);
+ s += pitch;
+ d9 = vld1_u8(s);
+ s += pitch;
+ d11 = vld1_u8(s);
+ s += pitch;
+ d13 = vld1_u8(s);
+ s += pitch;
+ d15 = vld1_u8(s);
+ s += pitch;
+ d17 = vld1_u8(s);
+ s += pitch;
+ d19 = vld1_u8(s);
+ s += pitch;
+ d21 = vld1_u8(s);
+
+ q3 = vcombine_u8(d6, d7);
+ q4 = vcombine_u8(d8, d9);
+ q5 = vcombine_u8(d10, d11);
+ q6 = vcombine_u8(d12, d13);
+ q7 = vcombine_u8(d14, d15);
+ q8 = vcombine_u8(d16, d17);
+ q9 = vcombine_u8(d18, d19);
+ q10 = vcombine_u8(d20, d21);
+
+ q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+ q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+ q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+ q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+ q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+ vreinterpretq_u16_u32(q2tmp2.val[0]));
+ q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+ vreinterpretq_u16_u32(q2tmp3.val[0]));
+ q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+ vreinterpretq_u16_u32(q2tmp2.val[1]));
+ q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+ vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+ q2tmp8 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+ vreinterpretq_u8_u16(q2tmp5.val[0]));
+ q2tmp9 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+ vreinterpretq_u8_u16(q2tmp5.val[1]));
+ q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+ vreinterpretq_u8_u16(q2tmp7.val[0]));
+ q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+ vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+ q3 = q2tmp8.val[0];
+ q4 = q2tmp8.val[1];
+ q5 = q2tmp9.val[0];
+ q6 = q2tmp9.val[1];
+ q7 = q2tmp10.val[0];
+ q8 = q2tmp10.val[1];
+ q9 = q2tmp11.val[0];
+ q10 = q2tmp11.val[1];
+
+ vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+ q5, q6, q7, q8, q9, q10,
+ &q5, &q6, &q7, &q8);
+
+ q4ResultL.val[0] = vget_low_u8(q5); // d10
+ q4ResultL.val[1] = vget_low_u8(q6); // d12
+ q4ResultL.val[2] = vget_low_u8(q7); // d14
+ q4ResultL.val[3] = vget_low_u8(q8); // d16
+ q4ResultH.val[0] = vget_high_u8(q5); // d11
+ q4ResultH.val[1] = vget_high_u8(q6); // d13
+ q4ResultH.val[2] = vget_high_u8(q7); // d15
+ q4ResultH.val[3] = vget_high_u8(q8); // d17
+
+ d = src - 2;
+ vst4_lane_u8(d, q4ResultL, 0);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 1);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 2);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 3);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 4);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 5);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 6);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultL, 7);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 0);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 1);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 2);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 3);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 4);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 5);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 6);
+ d += pitch;
+ vst4_lane_u8(d, q4ResultH, 7);
+ return;
+}
+
+void vp8_loop_filter_vertical_edge_uv_neon(
+ unsigned char *u,
+ int pitch,
+ unsigned char blimit,
+ unsigned char limit,
+ unsigned char thresh,
+ unsigned char *v) {
+ unsigned char *us, *ud;
+ unsigned char *vs, *vd;
+ uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+ uint8x16_t q5, q6, q7, q8, q9, q10;
+ uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+ uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+ uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+ uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+ uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+ uint8x8x4_t q4ResultH, q4ResultL;
+
+ qblimit = vdupq_n_u8(blimit);
+ qlimit = vdupq_n_u8(limit);
+ qthresh = vdupq_n_u8(thresh);
+
+ us = u - 4;
+ d6 = vld1_u8(us);
+ us += pitch;
+ d8 = vld1_u8(us);
+ us += pitch;
+ d10 = vld1_u8(us);
+ us += pitch;
+ d12 = vld1_u8(us);
+ us += pitch;
+ d14 = vld1_u8(us);
+ us += pitch;
+ d16 = vld1_u8(us);
+ us += pitch;
+ d18 = vld1_u8(us);
+ us += pitch;
+ d20 = vld1_u8(us);
+
+ vs = v - 4;
+ d7 = vld1_u8(vs);
+ vs += pitch;
+ d9 = vld1_u8(vs);
+ vs += pitch;
+ d11 = vld1_u8(vs);
+ vs += pitch;
+ d13 = vld1_u8(vs);
+ vs += pitch;
+ d15 = vld1_u8(vs);
+ vs += pitch;
+ d17 = vld1_u8(vs);
+ vs += pitch;
+ d19 = vld1_u8(vs);
+ vs += pitch;
+ d21 = vld1_u8(vs);
+
+ q3 = vcombine_u8(d6, d7);
+ q4 = vcombine_u8(d8, d9);
+ q5 = vcombine_u8(d10, d11);
+ q6 = vcombine_u8(d12, d13);
+ q7 = vcombine_u8(d14, d15);
+ q8 = vcombine_u8(d16, d17);
+ q9 = vcombine_u8(d18, d19);
+ q10 = vcombine_u8(d20, d21);
+
+ q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+ q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+ q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+ q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+ q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+ vreinterpretq_u16_u32(q2tmp2.val[0]));
+ q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+ vreinterpretq_u16_u32(q2tmp3.val[0]));
+ q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+ vreinterpretq_u16_u32(q2tmp2.val[1]));
+ q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+ vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+ q2tmp8 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+ vreinterpretq_u8_u16(q2tmp5.val[0]));
+ q2tmp9 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+ vreinterpretq_u8_u16(q2tmp5.val[1]));
+ q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+ vreinterpretq_u8_u16(q2tmp7.val[0]));
+ q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+ vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+ q3 = q2tmp8.val[0];
+ q4 = q2tmp8.val[1];
+ q5 = q2tmp9.val[0];
+ q6 = q2tmp9.val[1];
+ q7 = q2tmp10.val[0];
+ q8 = q2tmp10.val[1];
+ q9 = q2tmp11.val[0];
+ q10 = q2tmp11.val[1];
+
+ vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+ q5, q6, q7, q8, q9, q10,
+ &q5, &q6, &q7, &q8);
+
+ q4ResultL.val[0] = vget_low_u8(q5); // d10
+ q4ResultL.val[1] = vget_low_u8(q6); // d12
+ q4ResultL.val[2] = vget_low_u8(q7); // d14
+ q4ResultL.val[3] = vget_low_u8(q8); // d16
+ ud = u - 2;
+ vst4_lane_u8(ud, q4ResultL, 0);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 1);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 2);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 3);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 4);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 5);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 6);
+ ud += pitch;
+ vst4_lane_u8(ud, q4ResultL, 7);
+
+ q4ResultH.val[0] = vget_high_u8(q5); // d11
+ q4ResultH.val[1] = vget_high_u8(q6); // d13
+ q4ResultH.val[2] = vget_high_u8(q7); // d15
+ q4ResultH.val[3] = vget_high_u8(q8); // d17
+ vd = v - 2;
+ vst4_lane_u8(vd, q4ResultH, 0);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 1);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 2);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 3);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 4);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 5);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 6);
+ vd += pitch;
+ vst4_lane_u8(vd, q4ResultH, 7);
+ return;
+}
diff --git a/vp8/encoder/arm/neon/denoising_neon.c b/vp8/encoder/arm/neon/denoising_neon.c
index 3f8539759..23dc0a967 100644
--- a/vp8/encoder/arm/neon/denoising_neon.c
+++ b/vp8/encoder/arm/neon/denoising_neon.c
@@ -68,14 +68,11 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
int mc_running_avg_y_stride = mc_running_avg->y_stride;
unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
int running_avg_y_stride = running_avg->y_stride;
+ int64x2_t v_sum_diff_total = vdupq_n_s64(0);
/* Go over lines. */
int i;
- int sum_diff = 0;
for (i = 0; i < 16; ++i) {
- int8x16_t v_sum_diff = vdupq_n_s8(0);
- uint8x16_t v_running_avg_y;
-
/* Load inputs. */
const uint8x16_t v_sig = vld1q_u8(sig);
const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
@@ -117,12 +114,9 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
v_abs_adjustment);
const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
v_abs_adjustment);
- v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
+
+ uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
- v_sum_diff = vqaddq_s8(v_sum_diff,
- vreinterpretq_s8_u8(v_pos_adjustment));
- v_sum_diff = vqsubq_s8(v_sum_diff,
- vreinterpretq_s8_u8(v_neg_adjustment));
/* Store results. */
vst1q_u8(running_avg_y, v_running_avg_y);
@@ -131,23 +125,19 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
* for this macroblock.
*/
{
- int s0 = vgetq_lane_s8(v_sum_diff, 0) +
- vgetq_lane_s8(v_sum_diff, 1) +
- vgetq_lane_s8(v_sum_diff, 2) +
- vgetq_lane_s8(v_sum_diff, 3);
- int s1 = vgetq_lane_s8(v_sum_diff, 4) +
- vgetq_lane_s8(v_sum_diff, 5) +
- vgetq_lane_s8(v_sum_diff, 6) +
- vgetq_lane_s8(v_sum_diff, 7);
- int s2 = vgetq_lane_s8(v_sum_diff, 8) +
- vgetq_lane_s8(v_sum_diff, 9) +
- vgetq_lane_s8(v_sum_diff, 10) +
- vgetq_lane_s8(v_sum_diff, 11);
- int s3 = vgetq_lane_s8(v_sum_diff, 12) +
- vgetq_lane_s8(v_sum_diff, 13) +
- vgetq_lane_s8(v_sum_diff, 14) +
- vgetq_lane_s8(v_sum_diff, 15);
- sum_diff += s0 + s1+ s2 + s3;
+ const int8x16_t v_sum_diff =
+ vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+ vreinterpretq_s8_u8(v_neg_adjustment));
+
+ const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+
+ const int32x4_t fedc_ba98_7654_3210 =
+ vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+
+ const int64x2_t fedcba98_76543210 =
+ vpaddlq_s32(fedc_ba98_7654_3210);
+
+ v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
}
/* Update pointers for next iteration. */
@@ -157,11 +147,20 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
}
/* Too much adjustments => copy block. */
- if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
- return COPY_BLOCK;
+ {
+ const int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+ vget_low_s64(v_sum_diff_total));
+ const int s0 = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+ if (s0 > SUM_DIFF_THRESHOLD)
+ return COPY_BLOCK;
+ }
/* Tell above level that block was filtered. */
- vp8_copy_mem16x16(running_avg->y_buffer + y_offset, running_avg_y_stride,
- signal->thismb, sig_stride);
+ running_avg_y -= running_avg_y_stride * 16;
+ sig -= sig_stride * 16;
+
+ vp8_copy_mem16x16(running_avg_y, running_avg_y_stride, sig, sig_stride);
+
return FILTER_BLOCK;
}
diff --git a/vp8/encoder/x86/quantize_ssse3.asm b/vp8/encoder/x86/quantize_ssse3.asm
deleted file mode 100644
index 7b1dc119f..000000000
--- a/vp8/encoder/x86/quantize_ssse3.asm
+++ /dev/null
@@ -1,138 +0,0 @@
-;
-; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license and patent
-; grant that can be found in the LICENSE file in the root of the source
-; tree. All contributing project authors may be found in the AUTHORS
-; file in the root of the source tree.
-;
-
-
-%include "vpx_ports/x86_abi_support.asm"
-%include "vp8_asm_enc_offsets.asm"
-
-
-; void vp8_fast_quantize_b_ssse3 | arg
-; (BLOCK *b, | 0
-; BLOCKD *d) | 1
-;
-
-global sym(vp8_fast_quantize_b_ssse3) PRIVATE
-sym(vp8_fast_quantize_b_ssse3):
- push rbp
- mov rbp, rsp
- GET_GOT rbx
-
-%if ABI_IS_32BIT
- push rdi
- push rsi
-%else
- %if LIBVPX_YASM_WIN64
- push rdi
- push rsi
- %endif
-%endif
- ; end prolog
-
-%if ABI_IS_32BIT
- mov rdi, arg(0) ; BLOCK *b
- mov rsi, arg(1) ; BLOCKD *d
-%else
- %if LIBVPX_YASM_WIN64
- mov rdi, rcx ; BLOCK *b
- mov rsi, rdx ; BLOCKD *d
- %else
- ;mov rdi, rdi ; BLOCK *b
- ;mov rsi, rsi ; BLOCKD *d
- %endif
-%endif
-
- mov rax, [rdi + vp8_block_coeff]
- mov rcx, [rdi + vp8_block_round]
- mov rdx, [rdi + vp8_block_quant_fast]
-
- ; coeff
- movdqa xmm0, [rax]
- movdqa xmm4, [rax + 16]
-
- ; round
- movdqa xmm2, [rcx]
- movdqa xmm3, [rcx + 16]
-
- movdqa xmm1, xmm0
- movdqa xmm5, xmm4
-
- ; sz = z >> 15
- psraw xmm0, 15
- psraw xmm4, 15
-
- pabsw xmm1, xmm1
- pabsw xmm5, xmm5
-
- paddw xmm1, xmm2
- paddw xmm5, xmm3
-
- ; quant_fast
- pmulhw xmm1, [rdx]
- pmulhw xmm5, [rdx + 16]
-
- mov rax, [rsi + vp8_blockd_qcoeff]
- mov rdi, [rsi + vp8_blockd_dequant]
- mov rcx, [rsi + vp8_blockd_dqcoeff]
-
- movdqa xmm2, xmm1 ;store y for getting eob
- movdqa xmm3, xmm5
-
- pxor xmm1, xmm0
- pxor xmm5, xmm4
- psubw xmm1, xmm0
- psubw xmm5, xmm4
-
- movdqa [rax], xmm1
- movdqa [rax + 16], xmm5
-
- movdqa xmm0, [rdi]
- movdqa xmm4, [rdi + 16]
-
- pmullw xmm0, xmm1
- pmullw xmm4, xmm5
- pxor xmm1, xmm1
-
- pcmpgtw xmm2, xmm1 ;calculate eob
- pcmpgtw xmm3, xmm1
- packsswb xmm2, xmm3
- pshufb xmm2, [GLOBAL(zz_shuf)]
-
- pmovmskb edx, xmm2
-
- movdqa [rcx], xmm0 ;store dqcoeff
- movdqa [rcx + 16], xmm4 ;store dqcoeff
- mov rcx, [rsi + vp8_blockd_eob]
-
- bsr eax, edx ;count 0
- add eax, 1
-
- cmp edx, 0 ;if all 0, eob=0
- cmove eax, edx
-
- mov BYTE PTR [rcx], al ;store eob
-
- ; begin epilog
-%if ABI_IS_32BIT
- pop rsi
- pop rdi
-%else
- %if LIBVPX_YASM_WIN64
- pop rsi
- pop rdi
- %endif
-%endif
-
- RESTORE_GOT
- pop rbp
- ret
-
-SECTION_RODATA
-align 16
-zz_shuf:
- db 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
diff --git a/vp8/encoder/x86/quantize_ssse3.c b/vp8/encoder/x86/quantize_ssse3.c
new file mode 100644
index 000000000..9b4471d4f
--- /dev/null
+++ b/vp8/encoder/x86/quantize_ssse3.c
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <tmmintrin.h> /* SSSE3 */
+
+#include "vp8/encoder/block.h"
+
+/* bitscan reverse (bsr) */
+#if defined(_MSC_VER)
+#include <intrin.h>
+#pragma intrinsic(_BitScanReverse)
+static int bsr(int mask) {
+ int eob;
+ _BitScanReverse(&eob, mask);
+ eob++;
+ if (mask == 0)
+ eob = 0;
+ return eob;
+}
+#else
+static int bsr(int mask) {
+ int eob;
+ asm volatile("bsr %1, %0" : "=r" (eob) : "r" (mask) : "flags");
+ eob++;
+ if (mask == 0)
+ eob = 0;
+ return eob;
+}
+#endif
+
+void vp8_fast_quantize_b_ssse3(BLOCK *b, BLOCKD *d) {
+ int eob, mask;
+
+ __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+ __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
+ __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+ __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+ __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
+ __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
+ __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+ __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+
+ __m128i sz0, sz1, x, x0, x1, y0, y1, zeros, abs0, abs1;
+
+ DECLARE_ALIGNED(16, const uint8_t, pshufb_zig_zag_mask[16]) =
+ { 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15 };
+ __m128i zig_zag = _mm_load_si128((const __m128i *)pshufb_zig_zag_mask);
+
+ /* sign of z: z >> 15 */
+ sz0 = _mm_srai_epi16(z0, 15);
+ sz1 = _mm_srai_epi16(z1, 15);
+
+ /* x = abs(z) */
+ x0 = _mm_abs_epi16(z0);
+ x1 = _mm_abs_epi16(z1);
+
+ /* x += round */
+ x0 = _mm_add_epi16(x0, round0);
+ x1 = _mm_add_epi16(x1, round1);
+
+ /* y = (x * quant) >> 16 */
+ y0 = _mm_mulhi_epi16(x0, quant_fast0);
+ y1 = _mm_mulhi_epi16(x1, quant_fast1);
+
+ /* ASM saves Y for EOB */
+ /* I think we can ignore that because adding the sign doesn't change anything
+ * and multiplying 0 by dequant is OK as well */
+ abs0 = y0;
+ abs1 = y1;
+
+ /* Restore the sign bit. */
+ y0 = _mm_xor_si128(y0, sz0);
+ y1 = _mm_xor_si128(y1, sz1);
+ x0 = _mm_sub_epi16(y0, sz0);
+ x1 = _mm_sub_epi16(y1, sz1);
+
+ /* qcoeff = x */
+ _mm_store_si128((__m128i *)(d->qcoeff), x0);
+ _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
+
+ /* x * dequant */
+ x0 = _mm_mullo_epi16(x0, dequant0);
+ x1 = _mm_mullo_epi16(x1, dequant1);
+
+ /* dqcoeff = x * dequant */
+ _mm_store_si128((__m128i *)(d->dqcoeff), x0);
+ _mm_store_si128((__m128i *)(d->dqcoeff + 8), x1);
+
+ zeros = _mm_setzero_si128();
+
+ x0 = _mm_cmpgt_epi16(abs0, zeros);
+ x1 = _mm_cmpgt_epi16(abs1, zeros);
+
+ x = _mm_packs_epi16(x0, x1);
+
+ x = _mm_shuffle_epi8(x, zig_zag);
+
+ mask = _mm_movemask_epi8(x);
+
+ eob = bsr(mask);
+
+ *d->eob = 0xFF & eob;
+}
diff --git a/vp8/vp8_common.mk b/vp8/vp8_common.mk
index cdb271664..45063bd69 100644
--- a/vp8/vp8_common.mk
+++ b/vp8/vp8_common.mk
@@ -159,8 +159,6 @@ VP8_COMMON_SRCS-$(HAVE_MEDIA) += common/arm/armv6/vp8_variance_halfpixvar16x16_
VP8_COMMON_SRCS-$(HAVE_MEDIA) += common/arm/armv6/vp8_variance_halfpixvar16x16_hv_armv6$(ASM)
# common (neon)
-VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iwalsh_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/loopfilter_neon$(ASM)
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/loopfiltersimplehorizontaledge_neon$(ASM)
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/loopfiltersimpleverticaledge_neon$(ASM)
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/mbloopfilter_neon$(ASM)
@@ -186,6 +184,8 @@ VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/dc_only_idct_add_neon.c
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/dequant_idct_neon.c
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/dequantizeb_neon.c
VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/idct_dequant_full_2x_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iwalsh_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/loopfilter_neon.c
$(eval $(call rtcd_h_template,vp8_rtcd,vp8/common/rtcd_defs.pl))
diff --git a/vp8/vp8cx.mk b/vp8/vp8cx.mk
index d7c6dd1e1..607382b4c 100644
--- a/vp8/vp8cx.mk
+++ b/vp8/vp8cx.mk
@@ -88,6 +88,7 @@ VP8_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp8_enc_stubs_mmx.c
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/fwalsh_sse2.asm
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/quantize_sse2.c
+VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.c
ifeq ($(CONFIG_TEMPORAL_DENOISING),yes)
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/denoising_sse2.c
@@ -96,7 +97,6 @@ endif
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/subtract_sse2.asm
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp8_enc_stubs_sse2.c
-VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.asm
VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.asm
VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/quantize_mmx.asm
VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/encodeopt.asm
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index 6b9737076..364ea3a9e 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -276,6 +276,7 @@ typedef struct VP9EncoderConfig {
int arnr_max_frames;
int arnr_strength;
+ int arnr_type;
int tile_columns;
int tile_rows;
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index a2fc1bb77..f854356b9 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -3250,9 +3250,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
// an unfiltered alternative. We allow near/nearest as well
// because they may result in zero-zero MVs but be cheaper.
if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
- const int altref_zero_mask =
+ mode_skip_mask =
~((1 << THR_NEARESTA) | (1 << THR_NEARA) | (1 << THR_ZEROA));
- mode_skip_mask |= altref_zero_mask;
if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
mode_skip_mask |= (1 << THR_NEARA);
if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index a176bbf91..a5234cd9e 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -355,33 +355,74 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance) {
VP9_COMMON *const cm = &cpi->common;
int frame = 0;
+ int frames_to_blur_backward = 0;
+ int frames_to_blur_forward = 0;
int frames_to_blur = 0;
int start_frame = 0;
int strength = cpi->active_arnr_strength;
+ int blur_type = cpi->oxcf.arnr_type;
int max_frames = cpi->active_arnr_frames;
- int frames_to_blur_backward = distance;
- int frames_to_blur_forward = vp9_lookahead_depth(cpi->lookahead)
- - (distance + 1);
+ const int num_frames_backward = distance;
+ const int num_frames_forward = vp9_lookahead_depth(cpi->lookahead)
+ - (num_frames_backward + 1);
struct scale_factors sf;
- // Determine which input frames to filter.
- if (frames_to_blur_forward > frames_to_blur_backward)
- frames_to_blur_forward = frames_to_blur_backward;
+ switch (blur_type) {
+ case 1:
+ // Backward Blur
+ frames_to_blur_backward = num_frames_backward;
- if (frames_to_blur_backward > frames_to_blur_forward)
- frames_to_blur_backward = frames_to_blur_forward;
+ if (frames_to_blur_backward >= max_frames)
+ frames_to_blur_backward = max_frames - 1;
- // When max_frames is even we have 1 more frame backward than forward
- if (frames_to_blur_forward > (max_frames - 1) / 2)
- frames_to_blur_forward = (max_frames - 1) / 2;
+ frames_to_blur = frames_to_blur_backward + 1;
+ break;
- if (frames_to_blur_backward > (max_frames / 2))
- frames_to_blur_backward = max_frames / 2;
+ case 2:
+ // Forward Blur
+ frames_to_blur_forward = num_frames_forward;
- frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1;
+ if (frames_to_blur_forward >= max_frames)
+ frames_to_blur_forward = max_frames - 1;
+
+ frames_to_blur = frames_to_blur_forward + 1;
+ break;
+
+ case 3:
+ default:
+ // Center Blur
+ frames_to_blur_forward = num_frames_forward;
+ frames_to_blur_backward = num_frames_backward;
+
+ if (frames_to_blur_forward > frames_to_blur_backward)
+ frames_to_blur_forward = frames_to_blur_backward;
+
+ if (frames_to_blur_backward > frames_to_blur_forward)
+ frames_to_blur_backward = frames_to_blur_forward;
+
+ // When max_frames is even we have 1 more frame backward than forward
+ if (frames_to_blur_forward > (max_frames - 1) / 2)
+ frames_to_blur_forward = ((max_frames - 1) / 2);
+
+ if (frames_to_blur_backward > (max_frames / 2))
+ frames_to_blur_backward = (max_frames / 2);
+
+ frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1;
+ break;
+ }
start_frame = distance + frames_to_blur_forward;
+#ifdef DEBUGFWG
+ // DEBUG FWG
+ printf(
+ "max:%d FBCK:%d FFWD:%d ftb:%d ftbbck:%d ftbfwd:%d sei:%d lasei:%d "
+ "start:%d",
+ max_frames, num_frames_backward, num_frames_forward, frames_to_blur,
+ frames_to_blur_backward, frames_to_blur_forward, cpi->source_encode_index,
+ cpi->last_alt_ref_sei, start_frame);
+#endif
+
// Setup scaling factors. Scaling on each of the arnr frames is not supported
vp9_setup_scale_factors_for_frame(&sf,
get_frame_new_buffer(cm)->y_crop_width,
@@ -390,7 +431,7 @@ void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance) {
// Setup frame pointers, NULL indicates frame not included in filter
vp9_zero(cpi->frames);
- for (frame = 0; frame < frames_to_blur; ++frame) {
+ for (frame = 0; frame < frames_to_blur; frame++) {
int which_buffer = start_frame - frame;
struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead,
which_buffer);
@@ -404,11 +445,11 @@ void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance) {
void vp9_configure_arnr_filter(VP9_COMP *cpi,
const unsigned int frames_to_arnr,
const int group_boost) {
- int q;
int half_gf_int;
int frames_after_arf;
- int frames_bwd;
- int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+ int frames_bwd = cpi->oxcf.arnr_max_frames - 1;
+ int frames_fwd = cpi->oxcf.arnr_max_frames - 1;
+ int q;
// Define the arnr filter width for this group of frames. We only
// filter frames that lie within a distance of half the GF interval
@@ -420,26 +461,47 @@ void vp9_configure_arnr_filter(VP9_COMP *cpi,
frames_after_arf = vp9_lookahead_depth(cpi->lookahead)
- frames_to_arnr - 1;
- if (frames_fwd > frames_after_arf)
- frames_fwd = frames_after_arf;
- if (frames_fwd > half_gf_int)
- frames_fwd = half_gf_int;
-
- frames_bwd = frames_fwd;
-
- // For even length filter there is one more frame backward
- // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
- if (frames_bwd < half_gf_int)
- frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1;
+ switch (cpi->oxcf.arnr_type) {
+ case 1: // Backward filter
+ frames_fwd = 0;
+ if (frames_bwd > half_gf_int)
+ frames_bwd = half_gf_int;
+ break;
+
+ case 2: // Forward filter
+ if (frames_fwd > half_gf_int)
+ frames_fwd = half_gf_int;
+ if (frames_fwd > frames_after_arf)
+ frames_fwd = frames_after_arf;
+ frames_bwd = 0;
+ break;
+
+ case 3: // Centered filter
+ default:
+ frames_fwd >>= 1;
+ if (frames_fwd > frames_after_arf)
+ frames_fwd = frames_after_arf;
+ if (frames_fwd > half_gf_int)
+ frames_fwd = half_gf_int;
+
+ frames_bwd = frames_fwd;
+
+ // For even length filter there is one more frame backward
+ // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+ if (frames_bwd < half_gf_int)
+ frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1;
+ break;
+ }
cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd;
// Adjust the strength based on active max q
if (cpi->common.current_video_frame > 1)
- q = ((int)vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME]));
+ q = ((int)vp9_convert_qindex_to_q(
+ cpi->rc.avg_frame_qindex[INTER_FRAME]));
else
- q = ((int)vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[KEY_FRAME]));
-
+ q = ((int)vp9_convert_qindex_to_q(
+ cpi->rc.avg_frame_qindex[KEY_FRAME]));
if (q > 16) {
cpi->active_arnr_strength = cpi->oxcf.arnr_strength;
} else {
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index 1ca9fb93a..449e7d897 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -30,6 +30,7 @@ struct vp9_extracfg {
unsigned int tile_rows;
unsigned int arnr_max_frames;
unsigned int arnr_strength;
+ unsigned int arnr_type;
vp8e_tuning tuning;
unsigned int cq_level; // constrained quality level
unsigned int rc_max_intra_bitrate_pct;
@@ -59,6 +60,7 @@ static const struct extraconfig_map extracfg_map[] = {
0, // tile_rows
7, // arnr_max_frames
5, // arnr_strength
+ 3, // arnr_type
VP8_TUNE_PSNR, // tuning
10, // cq_level
0, // rc_max_intra_bitrate_pct
@@ -201,6 +203,7 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
RANGE_CHECK_HI(extra_cfg, sharpness, 7);
RANGE_CHECK(extra_cfg, arnr_max_frames, 0, 15);
RANGE_CHECK_HI(extra_cfg, arnr_strength, 6);
+ RANGE_CHECK(extra_cfg, arnr_type, 1, 3);
RANGE_CHECK(extra_cfg, cq_level, 0, 63);
// TODO(yaowu): remove this when ssim tuning is implemented for vp9
@@ -364,6 +367,7 @@ static vpx_codec_err_t set_encoder_config(
oxcf->arnr_max_frames = extra_cfg->arnr_max_frames;
oxcf->arnr_strength = extra_cfg->arnr_strength;
+ oxcf->arnr_type = extra_cfg->arnr_type;
oxcf->tuning = extra_cfg->tuning;
@@ -494,6 +498,7 @@ static vpx_codec_err_t ctrl_set_param(vpx_codec_alg_priv_t *ctx, int ctrl_id,
MAP(VP9E_SET_TILE_ROWS, extra_cfg.tile_rows);
MAP(VP8E_SET_ARNR_MAXFRAMES, extra_cfg.arnr_max_frames);
MAP(VP8E_SET_ARNR_STRENGTH, extra_cfg.arnr_strength);
+ MAP(VP8E_SET_ARNR_TYPE, extra_cfg.arnr_type);
MAP(VP8E_SET_TUNING, extra_cfg.tuning);
MAP(VP8E_SET_CQ_LEVEL, extra_cfg.cq_level);
MAP(VP8E_SET_MAX_INTRA_BITRATE_PCT, extra_cfg.rc_max_intra_bitrate_pct);
@@ -1107,6 +1112,7 @@ static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
{VP9E_SET_TILE_ROWS, ctrl_set_param},
{VP8E_SET_ARNR_MAXFRAMES, ctrl_set_param},
{VP8E_SET_ARNR_STRENGTH, ctrl_set_param},
+ {VP8E_SET_ARNR_TYPE, ctrl_set_param},
{VP8E_SET_TUNING, ctrl_set_param},
{VP8E_SET_CQ_LEVEL, ctrl_set_param},
{VP8E_SET_MAX_INTRA_BITRATE_PCT, ctrl_set_param},
diff --git a/vpx/vp8cx.h b/vpx/vp8cx.h
index 8944a2664..bc8df807e 100644
--- a/vpx/vp8cx.h
+++ b/vpx/vp8cx.h
@@ -160,8 +160,10 @@ enum vp8e_enc_control_id {
scale as used by the rc_*_quantizer config
parameters */
VP8E_SET_ARNR_MAXFRAMES, /**< control function to set the max number of frames blurred creating arf*/
- VP8E_SET_ARNR_STRENGTH, /**< control function to set the filter strength for the arf */
- VP8E_SET_ARNR_TYPE, /**< control function to set the type of filter to use for the arf*/
+ VP8E_SET_ARNR_STRENGTH, //!< control function to set the filter
+ //!< strength for the arf
+ VP8E_SET_ARNR_TYPE, //!< control function to set the type of
+ //!< filter to use for the arf
VP8E_SET_TUNING, /**< control function to set visual tuning */
/*!\brief control function to set constrained quality level
*
diff --git a/vpxenc.c b/vpxenc.c
index 3b8591e30..8e8ed2344 100644
--- a/vpxenc.c
+++ b/vpxenc.c
@@ -363,7 +363,7 @@ static const arg_def_t frame_periodic_boost = ARG_DEF(
static const arg_def_t *vp9_args[] = {
&cpu_used, &auto_altref, &noise_sens, &sharpness, &static_thresh,
- &tile_cols, &tile_rows, &arnr_maxframes, &arnr_strength,
+ &tile_cols, &tile_rows, &arnr_maxframes, &arnr_strength, &arnr_type,
&tune_ssim, &cq_level, &max_intra_rate_pct, &lossless,
&frame_parallel_decoding, &aq_mode, &frame_periodic_boost,
NULL
@@ -372,7 +372,7 @@ static const int vp9_arg_ctrl_map[] = {
VP8E_SET_CPUUSED, VP8E_SET_ENABLEAUTOALTREF,
VP8E_SET_NOISE_SENSITIVITY, VP8E_SET_SHARPNESS, VP8E_SET_STATIC_THRESHOLD,
VP9E_SET_TILE_COLUMNS, VP9E_SET_TILE_ROWS,
- VP8E_SET_ARNR_MAXFRAMES, VP8E_SET_ARNR_STRENGTH,
+ VP8E_SET_ARNR_MAXFRAMES, VP8E_SET_ARNR_STRENGTH, VP8E_SET_ARNR_TYPE,
VP8E_SET_TUNING, VP8E_SET_CQ_LEVEL, VP8E_SET_MAX_INTRA_BITRATE_PCT,
VP9E_SET_LOSSLESS, VP9E_SET_FRAME_PARALLEL_DECODING, VP9E_SET_AQ_MODE,
VP9E_SET_FRAME_PERIODIC_BOOST,
generated by cgit v1.2.3 (git 2.25.1) at 2024-10-14 23:34:40 +0000