void vp9_highbd_iht8x8_64_add_neon(const tran_low_t *input, uint16_t *dest,
int stride, int tx_type, int bd) {
int32x4_t a[16];
int16x8_t c[8];
a[0] = vld1q_s32(input);
a[1] = vld1q_s32(input + 4);
a[2] = vld1q_s32(input + 8);
a[3] = vld1q_s32(input + 12);
a[4] = vld1q_s32(input + 16);
a[5] = vld1q_s32(input + 20);
a[6] = vld1q_s32(input + 24);
a[7] = vld1q_s32(input + 28);
a[8] = vld1q_s32(input + 32);
a[9] = vld1q_s32(input + 36);
a[10] = vld1q_s32(input + 40);
a[11] = vld1q_s32(input + 44);
a[12] = vld1q_s32(input + 48);
a[13] = vld1q_s32(input + 52);
a[14] = vld1q_s32(input + 56);
a[15] = vld1q_s32(input + 60);
if (bd == 8) {
c[0] = vcombine_s16(vmovn_s32(a[0]), vmovn_s32(a[1]));
c[1] = vcombine_s16(vmovn_s32(a[2]), vmovn_s32(a[3]));
c[2] = vcombine_s16(vmovn_s32(a[4]), vmovn_s32(a[5]));
c[3] = vcombine_s16(vmovn_s32(a[6]), vmovn_s32(a[7]));
c[4] = vcombine_s16(vmovn_s32(a[8]), vmovn_s32(a[9]));
c[5] = vcombine_s16(vmovn_s32(a[10]), vmovn_s32(a[11]));
c[6] = vcombine_s16(vmovn_s32(a[12]), vmovn_s32(a[13]));
c[7] = vcombine_s16(vmovn_s32(a[14]), vmovn_s32(a[15]));
switch (tx_type) {
case DCT_DCT: {
const int16x8_t cospis = vld1q_s16(kCospi);
const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24
const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28
idct8x8_64_1d_bd8(cospis0, cospis1, c);
idct8x8_64_1d_bd8(cospis0, cospis1, c);
break;
}
case ADST_DCT: {
const int16x8_t cospis = vld1q_s16(kCospi);
const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24
const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28
idct8x8_64_1d_bd8(cospis0, cospis1, c);
transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6],
&c[7]);
iadst8(c);
break;
}
case DCT_ADST: {
const int16x8_t cospis = vld1q_s16(kCospi);
const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24
const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28
transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6],
&c[7]);
iadst8(c);
idct8x8_64_1d_bd8(cospis0, cospis1, c);
break;
}
default: {
transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6],
&c[7]);
iadst8(c);
transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6],
&c[7]);
iadst8(c);
break;
}
}
c[0] = vrshrq_n_s16(c[0], 5);
c[1] = vrshrq_n_s16(c[1], 5);
c[2] = vrshrq_n_s16(c[2], 5);
c[3] = vrshrq_n_s16(c[3], 5);
c[4] = vrshrq_n_s16(c[4], 5);
c[5] = vrshrq_n_s16(c[5], 5);
c[6] = vrshrq_n_s16(c[6], 5);
c[7] = vrshrq_n_s16(c[7], 5);
} else {
switch (tx_type) {
case DCT_DCT: {
const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24
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]);
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 {
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]);
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;
}
case ADST_DCT: {
const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24
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],
&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;
}
case DCT_ADST: {
const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24
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],
&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]);
}
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],
&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],
&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;
}
}
c[0] = vcombine_s16(vrshrn_n_s32(a[0], 5), vrshrn_n_s32(a[4], 5));
c[1] = vcombine_s16(vrshrn_n_s32(a[8], 5), vrshrn_n_s32(a[12], 5));
c[2] = vcombine_s16(vrshrn_n_s32(a[1], 5), vrshrn_n_s32(a[5], 5));
c[3] = vcombine_s16(vrshrn_n_s32(a[9], 5), vrshrn_n_s32(a[13], 5));
c[4] = vcombine_s16(vrshrn_n_s32(a[2], 5), vrshrn_n_s32(a[6], 5));
c[5] = vcombine_s16(vrshrn_n_s32(a[10], 5), vrshrn_n_s32(a[14], 5));
c[6] = vcombine_s16(vrshrn_n_s32(a[3], 5), vrshrn_n_s32(a[7], 5));
c[7] = vcombine_s16(vrshrn_n_s32(a[11], 5), vrshrn_n_s32(a[15], 5));
}
highbd_add8x8(c, dest, stride, bd);
}
// vp9/common/vp9_scan.c:vp9_default_iscan_32x32 arranges the first 34 non-zero
// coefficients as follows:
// 0 1 2 3 4 5 6 7
// 0 0 2 5 10 17 25
// 1 1 4 8 15 22 30
// 2 3 7 12 18 28
// 3 6 11 16 23 31
// 4 9 14 19 29
// 5 13 20 26
// 6 21 27 33
// 7 24 32
void vpx_idct32_6_neon(const tran_low_t *input, int16_t *output) {
int16x8_t in[8], s1[32], s2[32], s3[32];
in[0] = load_tran_low_to_s16q(input);
input += 32;
in[1] = load_tran_low_to_s16q(input);
input += 32;
in[2] = load_tran_low_to_s16q(input);
input += 32;
in[3] = load_tran_low_to_s16q(input);
input += 32;
in[4] = load_tran_low_to_s16q(input);
input += 32;
in[5] = load_tran_low_to_s16q(input);
input += 32;
in[6] = load_tran_low_to_s16q(input);
input += 32;
in[7] = load_tran_low_to_s16q(input);
transpose_s16_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
&in[7]);
// stage 1
// input[1] * cospi_31_64 - input[31] * cospi_1_64 (but input[31] == 0)
s1[16] = multiply_shift_and_narrow_s16(in[1], cospi_31_64);
// input[1] * cospi_1_64 + input[31] * cospi_31_64 (but input[31] == 0)
s1[31] = multiply_shift_and_narrow_s16(in[1], cospi_1_64);
s1[20] = multiply_shift_and_narrow_s16(in[5], cospi_27_64);
s1[27] = multiply_shift_and_narrow_s16(in[5], cospi_5_64);
s1[23] = multiply_shift_and_narrow_s16(in[3], -cospi_29_64);
s1[24] = multiply_shift_and_narrow_s16(in[3], cospi_3_64);
// stage 2
s2[8] = multiply_shift_and_narrow_s16(in[2], cospi_30_64);
s2[15] = multiply_shift_and_narrow_s16(in[2], cospi_2_64);
// stage 3
s1[4] = multiply_shift_and_narrow_s16(in[4], cospi_28_64);
s1[7] = multiply_shift_and_narrow_s16(in[4], cospi_4_64);
s1[17] = multiply_accumulate_shift_and_narrow_s16(s1[16], -cospi_4_64, s1[31],
cospi_28_64);
s1[30] = multiply_accumulate_shift_and_narrow_s16(s1[16], cospi_28_64, s1[31],
cospi_4_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s1[20], -cospi_20_64,
s1[27], cospi_12_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s1[20], cospi_12_64, s1[27],
cospi_20_64);
s1[22] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_12_64,
s1[24], -cospi_20_64);
s1[25] = multiply_accumulate_shift_and_narrow_s16(s1[23], -cospi_20_64,
s1[24], cospi_12_64);
// stage 4
s1[0] = multiply_shift_and_narrow_s16(in[0], cospi_16_64);
s2[9] = multiply_accumulate_shift_and_narrow_s16(s2[8], -cospi_8_64, s2[15],
cospi_24_64);
s2[14] = multiply_accumulate_shift_and_narrow_s16(s2[8], cospi_24_64, s2[15],
cospi_8_64);
s2[20] = vsubq_s16(s1[23], s1[20]);
s2[21] = vsubq_s16(s1[22], s1[21]);
s2[22] = vaddq_s16(s1[21], s1[22]);
s2[23] = vaddq_s16(s1[20], s1[23]);
s2[24] = vaddq_s16(s1[24], s1[27]);
s2[25] = vaddq_s16(s1[25], s1[26]);
s2[26] = vsubq_s16(s1[25], s1[26]);
s2[27] = vsubq_s16(s1[24], s1[27]);
// stage 5
s1[5] = sub_multiply_shift_and_narrow_s16(s1[7], s1[4], cospi_16_64);
s1[6] = add_multiply_shift_and_narrow_s16(s1[4], s1[7], cospi_16_64);
s1[18] = multiply_accumulate_shift_and_narrow_s16(s1[17], -cospi_8_64, s1[30],
cospi_24_64);
s1[29] = multiply_accumulate_shift_and_narrow_s16(s1[17], cospi_24_64, s1[30],
cospi_8_64);
s1[19] = multiply_accumulate_shift_and_narrow_s16(s1[16], -cospi_8_64, s1[31],
cospi_24_64);
s1[28] = multiply_accumulate_shift_and_narrow_s16(s1[16], cospi_24_64, s1[31],
cospi_8_64);
s1[20] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_24_64,
s2[27], -cospi_8_64);
s1[27] = multiply_accumulate_shift_and_narrow_s16(s2[20], -cospi_8_64, s2[27],
cospi_24_64);
s1[21] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_24_64,
s2[26], -cospi_8_64);
s1[26] = multiply_accumulate_shift_and_narrow_s16(s2[21], -cospi_8_64, s2[26],
cospi_24_64);
// stage 6
s2[0] = vaddq_s16(s1[0], s1[7]);
s2[1] = vaddq_s16(s1[0], s1[6]);
s2[2] = vaddq_s16(s1[0], s1[5]);
s2[3] = vaddq_s16(s1[0], s1[4]);
s2[4] = vsubq_s16(s1[0], s1[4]);
s2[5] = vsubq_s16(s1[0], s1[5]);
s2[6] = vsubq_s16(s1[0], s1[6]);
s2[7] = vsubq_s16(s1[0], s1[7]);
s2[10] = sub_multiply_shift_and_narrow_s16(s2[14], s2[9], cospi_16_64);
s2[13] = add_multiply_shift_and_narrow_s16(s2[9], s2[14], cospi_16_64);
s2[11] = sub_multiply_shift_and_narrow_s16(s2[15], s2[8], cospi_16_64);
s2[12] = add_multiply_shift_and_narrow_s16(s2[8], s2[15], cospi_16_64);
s2[16] = vaddq_s16(s1[16], s2[23]);
s2[17] = vaddq_s16(s1[17], s2[22]);
s2[18] = vaddq_s16(s1[18], s1[21]);
s2[19] = vaddq_s16(s1[19], s1[20]);
s2[20] = vsubq_s16(s1[19], s1[20]);
s2[21] = vsubq_s16(s1[18], s1[21]);
s2[22] = vsubq_s16(s1[17], s2[22]);
s2[23] = vsubq_s16(s1[16], s2[23]);
s3[24] = vsubq_s16(s1[31], s2[24]);
s3[25] = vsubq_s16(s1[30], s2[25]);
s3[26] = vsubq_s16(s1[29], s1[26]);
s3[27] = vsubq_s16(s1[28], s1[27]);
s2[28] = vaddq_s16(s1[27], s1[28]);
s2[29] = vaddq_s16(s1[26], s1[29]);
s2[30] = vaddq_s16(s2[25], s1[30]);
s2[31] = vaddq_s16(s2[24], s1[31]);
// stage 7
s1[0] = vaddq_s16(s2[0], s2[15]);
s1[1] = vaddq_s16(s2[1], s2[14]);
s1[2] = vaddq_s16(s2[2], s2[13]);
s1[3] = vaddq_s16(s2[3], s2[12]);
s1[4] = vaddq_s16(s2[4], s2[11]);
s1[5] = vaddq_s16(s2[5], s2[10]);
s1[6] = vaddq_s16(s2[6], s2[9]);
s1[7] = vaddq_s16(s2[7], s2[8]);
s1[8] = vsubq_s16(s2[7], s2[8]);
s1[9] = vsubq_s16(s2[6], s2[9]);
s1[10] = vsubq_s16(s2[5], s2[10]);
s1[11] = vsubq_s16(s2[4], s2[11]);
s1[12] = vsubq_s16(s2[3], s2[12]);
s1[13] = vsubq_s16(s2[2], s2[13]);
s1[14] = vsubq_s16(s2[1], s2[14]);
s1[15] = vsubq_s16(s2[0], s2[15]);
s1[20] = sub_multiply_shift_and_narrow_s16(s3[27], s2[20], cospi_16_64);
s1[27] = add_multiply_shift_and_narrow_s16(s2[20], s3[27], cospi_16_64);
s1[21] = sub_multiply_shift_and_narrow_s16(s3[26], s2[21], cospi_16_64);
s1[26] = add_multiply_shift_and_narrow_s16(s2[21], s3[26], cospi_16_64);
s1[22] = sub_multiply_shift_and_narrow_s16(s3[25], s2[22], cospi_16_64);
s1[25] = add_multiply_shift_and_narrow_s16(s2[22], s3[25], cospi_16_64);
s1[23] = sub_multiply_shift_and_narrow_s16(s3[24], s2[23], cospi_16_64);
s1[24] = add_multiply_shift_and_narrow_s16(s2[23], s3[24], cospi_16_64);
// final stage
vst1q_s16(output, vaddq_s16(s1[0], s2[31]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[1], s2[30]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[2], s2[29]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[3], s2[28]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[4], s1[27]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[5], s1[26]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[6], s1[25]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[7], s1[24]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[8], s1[23]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[9], s1[22]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[10], s1[21]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[11], s1[20]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[12], s2[19]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[13], s2[18]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[14], s2[17]));
output += 8;
vst1q_s16(output, vaddq_s16(s1[15], s2[16]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[15], s2[16]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[14], s2[17]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[13], s2[18]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[12], s2[19]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[11], s1[20]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[10], s1[21]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[9], s1[22]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[8], s1[23]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[7], s1[24]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[6], s1[25]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[5], s1[26]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[4], s1[27]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[3], s2[28]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[2], s2[29]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[1], s2[30]));
output += 8;
vst1q_s16(output, vsubq_s16(s1[0], s2[31]));
}
