Skip to content

Commit

Permalink
Merge changes I9e0bf2c7,I695b4090
Browse files Browse the repository at this point in the history 
* changes:
  Fix a bug in vp9_highbd_iht8x8_64_add_neon
  Fix a bug in vp9_highbd_iht4x4_16_add_neon()
  • Loading branch information
@linfengz
linfengz authored and Gerrit Code Review committed Mar 15, 2018
2 parents 881c8ec + d8424d2 commit ddb3d7a
Show file tree
Hide file tree
Showing 4 changed files with 86 additions and 188 deletions.
4 changes: 0 additions & 4 deletions test/dct_test.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -629,14 +629,10 @@ INSTANTIATE_TEST_CASE_P(

static const FuncInfo ht_neon_func_info[] = {
#if CONFIG_VP9_HIGHBITDEPTH
// TODO(linfengz): reenable these functions once test vector failures are
// addressed.
#if 0
{ &vp9_highbd_fht4x4_c, &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4,
2 },
{ &vp9_highbd_fht8x8_c, &highbd_iht_wrapper<vp9_highbd_iht8x8_64_add_neon>, 8,
2 },
#endif
#endif
{ &vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1 },
{ &vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1 },
Expand Down
75 changes: 48 additions & 27 deletions vp9/common/arm/neon/vp9_highbd_iht4x4_add_neon.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,34 +23,55 @@
static INLINE void highbd_iadst4(int32x4_t *const io) {
const int32_t sinpis[4] = { sinpi_1_9, sinpi_2_9, sinpi_3_9, sinpi_4_9 };
const int32x4_t sinpi = vld1q_s32(sinpis);
int32x4_t s[8];

s[0] = vmulq_lane_s32(io[0], vget_low_s32(sinpi), 0);
s[1] = vmulq_lane_s32(io[0], vget_low_s32(sinpi), 1);
s[2] = vmulq_lane_s32(io[1], vget_high_s32(sinpi), 0);
s[3] = vmulq_lane_s32(io[2], vget_high_s32(sinpi), 1);
s[4] = vmulq_lane_s32(io[2], vget_low_s32(sinpi), 0);
s[5] = vmulq_lane_s32(io[3], vget_low_s32(sinpi), 1);
s[6] = vmulq_lane_s32(io[3], vget_high_s32(sinpi), 1);
s[7] = vsubq_s32(io[0], io[2]);
s[7] = vaddq_s32(s[7], io[3]);

s[0] = vaddq_s32(s[0], s[3]);
s[0] = vaddq_s32(s[0], s[5]);
s[1] = vsubq_s32(s[1], s[4]);
s[1] = vsubq_s32(s[1], s[6]);
int64x2x2_t s[7], t[4];
int32x4_t s7;

s[0].val[0] = vmull_lane_s32(vget_low_s32(io[0]), vget_low_s32(sinpi), 0);
s[0].val[1] = vmull_lane_s32(vget_high_s32(io[0]), vget_low_s32(sinpi), 0);
s[1].val[0] = vmull_lane_s32(vget_low_s32(io[0]), vget_low_s32(sinpi), 1);
s[1].val[1] = vmull_lane_s32(vget_high_s32(io[0]), vget_low_s32(sinpi), 1);
s[2].val[0] = vmull_lane_s32(vget_low_s32(io[1]), vget_high_s32(sinpi), 0);
s[2].val[1] = vmull_lane_s32(vget_high_s32(io[1]), vget_high_s32(sinpi), 0);
s[3].val[0] = vmull_lane_s32(vget_low_s32(io[2]), vget_high_s32(sinpi), 1);
s[3].val[1] = vmull_lane_s32(vget_high_s32(io[2]), vget_high_s32(sinpi), 1);
s[4].val[0] = vmull_lane_s32(vget_low_s32(io[2]), vget_low_s32(sinpi), 0);
s[4].val[1] = vmull_lane_s32(vget_high_s32(io[2]), vget_low_s32(sinpi), 0);
s[5].val[0] = vmull_lane_s32(vget_low_s32(io[3]), vget_low_s32(sinpi), 1);
s[5].val[1] = vmull_lane_s32(vget_high_s32(io[3]), vget_low_s32(sinpi), 1);
s[6].val[0] = vmull_lane_s32(vget_low_s32(io[3]), vget_high_s32(sinpi), 1);
s[6].val[1] = vmull_lane_s32(vget_high_s32(io[3]), vget_high_s32(sinpi), 1);
s7 = vsubq_s32(io[0], io[2]);
s7 = vaddq_s32(s7, io[3]);

s[0].val[0] = vaddq_s64(s[0].val[0], s[3].val[0]);
s[0].val[1] = vaddq_s64(s[0].val[1], s[3].val[1]);
s[0].val[0] = vaddq_s64(s[0].val[0], s[5].val[0]);
s[0].val[1] = vaddq_s64(s[0].val[1], s[5].val[1]);
s[1].val[0] = vsubq_s64(s[1].val[0], s[4].val[0]);
s[1].val[1] = vsubq_s64(s[1].val[1], s[4].val[1]);
s[1].val[0] = vsubq_s64(s[1].val[0], s[6].val[0]);
s[1].val[1] = vsubq_s64(s[1].val[1], s[6].val[1]);
s[3] = s[2];
s[2] = vmulq_lane_s32(s[7], vget_high_s32(sinpi), 0);

io[0] = vaddq_s32(s[0], s[3]);
io[1] = vaddq_s32(s[1], s[3]);
io[2] = s[2];
io[3] = vaddq_s32(s[0], s[1]);
io[3] = vsubq_s32(io[3], s[3]);
io[0] = vrshrq_n_s32(io[0], DCT_CONST_BITS);
io[1] = vrshrq_n_s32(io[1], DCT_CONST_BITS);
io[2] = vrshrq_n_s32(io[2], DCT_CONST_BITS);
io[3] = vrshrq_n_s32(io[3], DCT_CONST_BITS);
s[2].val[0] = vmull_lane_s32(vget_low_s32(s7), vget_high_s32(sinpi), 0);
s[2].val[1] = vmull_lane_s32(vget_high_s32(s7), vget_high_s32(sinpi), 0);

t[0].val[0] = vaddq_s64(s[0].val[0], s[3].val[0]);
t[0].val[1] = vaddq_s64(s[0].val[1], s[3].val[1]);
t[1].val[0] = vaddq_s64(s[1].val[0], s[3].val[0]);
t[1].val[1] = vaddq_s64(s[1].val[1], s[3].val[1]);
t[2] = s[2];
t[3].val[0] = vaddq_s64(s[0].val[0], s[1].val[0]);
t[3].val[1] = vaddq_s64(s[0].val[1], s[1].val[1]);
t[3].val[0] = vsubq_s64(t[3].val[0], s[3].val[0]);
t[3].val[1] = vsubq_s64(t[3].val[1], s[3].val[1]);
io[0] = vcombine_s32(vrshrn_n_s64(t[0].val[0], DCT_CONST_BITS),
vrshrn_n_s64(t[0].val[1], DCT_CONST_BITS));
io[1] = vcombine_s32(vrshrn_n_s64(t[1].val[0], DCT_CONST_BITS),
vrshrn_n_s64(t[1].val[1], DCT_CONST_BITS));
io[2] = vcombine_s32(vrshrn_n_s64(t[2].val[0], DCT_CONST_BITS),
vrshrn_n_s64(t[2].val[1], DCT_CONST_BITS));
io[3] = vcombine_s32(vrshrn_n_s64(t[3].val[0], DCT_CONST_BITS),
vrshrn_n_s64(t[3].val[1], DCT_CONST_BITS));
}

void vp9_highbd_iht4x4_16_add_neon(const tran_low_t *input, uint16_t *dest,
Expand Down
191 changes: 36 additions & 155 deletions vp9/common/arm/neon/vp9_highbd_iht8x8_add_neon.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -132,84 +132,10 @@ static INLINE int32x4_t sub_dct_const_round_shift_low_8_bd12(
return vcombine_s32(out_lo, out_hi);
}

static INLINE void iadst8_bd10(int32x4_t *const io0, int32x4_t *const io1,
int32x4_t *const io2, int32x4_t *const io3,
int32x4_t *const io4, int32x4_t *const io5,
int32x4_t *const io6, int32x4_t *const io7) {
const int32x4_t c0 =
create_s32x4_neon(cospi_2_64, cospi_30_64, cospi_10_64, cospi_22_64);
const int32x4_t c1 =
create_s32x4_neon(cospi_18_64, cospi_14_64, cospi_26_64, cospi_6_64);
const int32x4_t c2 =
create_s32x4_neon(cospi_16_64, 0, cospi_8_64, cospi_24_64);
int32x4_t x[8], t[4];
int32x4_t s[8];

x[0] = *io7;
x[1] = *io0;
x[2] = *io5;
x[3] = *io2;
x[4] = *io3;
x[5] = *io4;
x[6] = *io1;
x[7] = *io6;

// stage 1
iadst_butterfly_lane_0_1_bd10_neon(x[0], x[1], vget_low_s32(c0), &s[0],
&s[1]);
iadst_butterfly_lane_0_1_bd10_neon(x[2], x[3], vget_high_s32(c0), &s[2],
&s[3]);
iadst_butterfly_lane_0_1_bd10_neon(x[4], x[5], vget_low_s32(c1), &s[4],
&s[5]);
iadst_butterfly_lane_0_1_bd10_neon(x[6], x[7], vget_high_s32(c1), &s[6],
&s[7]);

x[0] = add_dct_const_round_shift_low_8_bd10(s[0], s[4]);
x[1] = add_dct_const_round_shift_low_8_bd10(s[1], s[5]);
x[2] = add_dct_const_round_shift_low_8_bd10(s[2], s[6]);
x[3] = add_dct_const_round_shift_low_8_bd10(s[3], s[7]);
x[4] = sub_dct_const_round_shift_low_8_bd10(s[0], s[4]);
x[5] = sub_dct_const_round_shift_low_8_bd10(s[1], s[5]);
x[6] = sub_dct_const_round_shift_low_8_bd10(s[2], s[6]);
x[7] = sub_dct_const_round_shift_low_8_bd10(s[3], s[7]);

// stage 2
t[0] = x[0];
t[1] = x[1];
t[2] = x[2];
t[3] = x[3];
iadst_butterfly_lane_0_1_bd10_neon(x[4], x[5], vget_high_s32(c2), &s[4],
&s[5]);
iadst_butterfly_lane_1_0_bd10_neon(x[7], x[6], vget_high_s32(c2), &s[7],
&s[6]);

x[0] = vaddq_s32(t[0], t[2]);
x[1] = vaddq_s32(t[1], t[3]);
x[2] = vsubq_s32(t[0], t[2]);
x[3] = vsubq_s32(t[1], t[3]);
x[4] = add_dct_const_round_shift_low_8_bd10(s[4], s[6]);
x[5] = add_dct_const_round_shift_low_8_bd10(s[5], s[7]);
x[6] = sub_dct_const_round_shift_low_8_bd10(s[4], s[6]);
x[7] = sub_dct_const_round_shift_low_8_bd10(s[5], s[7]);

// stage 3
iadst_half_butterfly_bd10_neon(x + 2, vget_low_s32(c2));
iadst_half_butterfly_bd10_neon(x + 6, vget_low_s32(c2));

*io0 = x[0];
*io1 = vnegq_s32(x[4]);
*io2 = x[6];
*io3 = vnegq_s32(x[2]);
*io4 = x[3];
*io5 = vnegq_s32(x[7]);
*io6 = x[5];
*io7 = vnegq_s32(x[1]);
}

static INLINE void iadst8_bd12(int32x4_t *const io0, int32x4_t *const io1,
int32x4_t *const io2, int32x4_t *const io3,
int32x4_t *const io4, int32x4_t *const io5,
int32x4_t *const io6, int32x4_t *const io7) {
static INLINE void highbd_iadst8(int32x4_t *const io0, int32x4_t *const io1,
int32x4_t *const io2, int32x4_t *const io3,
int32x4_t *const io4, int32x4_t *const io5,
int32x4_t *const io6, int32x4_t *const io7) {
const int32x4_t c0 =
create_s32x4_neon(cospi_2_64, cospi_30_64, cospi_10_64, cospi_22_64);
const int32x4_t c1 =
Expand Down Expand Up @@ -394,31 +320,17 @@ void vp9_highbd_iht8x8_64_add_neon(const tran_low_t *input, uint16_t *dest,
const int32x4_t cospis1 =
vld1q_s32(kCospi32 + 4); // cospi 4, 12, 20, 28

if (bd == 10) {
idct8x8_64_half1d_bd10(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3],
&a[4], &a[5], &a[6], &a[7]);
idct8x8_64_half1d_bd10(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11],
&a[12], &a[13], &a[14], &a[15]);
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
iadst8_bd10(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
iadst8_bd10(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3],
&a[4], &a[5], &a[6], &a[7]);
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11],
&a[12], &a[13], &a[14], &a[15]);
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
highbd_iadst8(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
highbd_iadst8(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
} else {
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3],
&a[4], &a[5], &a[6], &a[7]);
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11],
&a[12], &a[13], &a[14], &a[15]);
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
iadst8_bd12(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
iadst8_bd12(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
}
break;
}

Expand All @@ -427,67 +339,36 @@ void vp9_highbd_iht8x8_64_add_neon(const tran_low_t *input, uint16_t *dest,
const int32x4_t cospis1 =
vld1q_s32(kCospi32 + 4); // cospi 4, 12, 20, 28

if (bd == 10) {
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
iadst8_bd10(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13],
&a[14], &a[15]);
iadst8_bd10(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
highbd_iadst8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
highbd_iadst8(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
idct8x8_64_half1d_bd10(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9],
&a[2], &a[10], &a[3], &a[11]);
idct8x8_64_half1d_bd10(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13],
&a[6], &a[14], &a[7], &a[15]);
} else {
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
iadst8_bd12(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13],
&a[14], &a[15]);
iadst8_bd12(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9],
&a[2], &a[10], &a[3], &a[11]);
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13],
&a[6], &a[14], &a[7], &a[15]);
}
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9],
&a[2], &a[10], &a[3], &a[11]);
idct8x8_64_half1d_bd12(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13],
&a[6], &a[14], &a[7], &a[15]);
break;
}

default: {
assert(tx_type == ADST_ADST);
if (bd == 10) {
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
iadst8_bd10(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13],
&a[14], &a[15]);
iadst8_bd10(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
highbd_iadst8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
highbd_iadst8(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
iadst8_bd10(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
iadst8_bd10(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
highbd_iadst8(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
highbd_iadst8(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
} else {
transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6],
&a[7]);
iadst8_bd12(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]);
transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13],
&a[14], &a[15]);
iadst8_bd12(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14],
&a[15]);
transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3],
&a[11]);
iadst8_bd12(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]);
transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
iadst8_bd12(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7],
&a[15]);
}
break;
}
}
Expand Down
4 changes: 2 additions & 2 deletions vp9/common/vp9_rtcd_defs.pl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -103,8 +103,8 @@ ()
add_proto qw/void vp9_highbd_iht16x16_256_add/, "const tran_low_t *input, uint16_t *output, int pitch, int tx_type, int bd";

if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
specialize qw/vp9_highbd_iht4x4_16_add sse4_1/;
specialize qw/vp9_highbd_iht8x8_64_add sse4_1/;
specialize qw/vp9_highbd_iht4x4_16_add neon sse4_1/;
specialize qw/vp9_highbd_iht8x8_64_add neon sse4_1/;
specialize qw/vp9_highbd_iht16x16_256_add sse4_1/;
}
}
Expand Down

0 comments on commit ddb3d7a

Please sign in to comment.