Merge "VP8 for ARMv8 by using NEON intrinsics 07"

vp8/common/arm/neon/iwalsh_neon.asm

-;
-;  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

vp8/common/arm/neon/iwalsh_neon.c

+/*
+ *  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;
+}

vp8/vp8_common.mk

@@ -159,7 +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)
@@ -186,6 +185,7 @@ 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
 
 
 $(eval $(call rtcd_h_template,vp8_rtcd,vp8/common/rtcd_defs.pl))