static INLINE void TRANSPOSE8X8(
int16x8_t *q8s16,
int16x8_t *q9s16,
int16x8_t *q10s16,
int16x8_t *q11s16,
int16x8_t *q12s16,
int16x8_t *q13s16,
int16x8_t *q14s16,
int16x8_t *q15s16) {
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
d20s16 = vget_low_s16(*q10s16);
d21s16 = vget_high_s16(*q10s16);
d22s16 = vget_low_s16(*q11s16);
d23s16 = vget_high_s16(*q11s16);
d24s16 = vget_low_s16(*q12s16);
d25s16 = vget_high_s16(*q12s16);
d26s16 = vget_low_s16(*q13s16);
d27s16 = vget_high_s16(*q13s16);
d28s16 = vget_low_s16(*q14s16);
d29s16 = vget_high_s16(*q14s16);
d30s16 = vget_low_s16(*q15s16);
d31s16 = vget_high_s16(*q15s16);
*q8s16 = vcombine_s16(d16s16, d24s16); // vswp d17, d24
*q9s16 = vcombine_s16(d18s16, d26s16); // vswp d19, d26
*q10s16 = vcombine_s16(d20s16, d28s16); // vswp d21, d28
*q11s16 = vcombine_s16(d22s16, d30s16); // vswp d23, d30
*q12s16 = vcombine_s16(d17s16, d25s16);
*q13s16 = vcombine_s16(d19s16, d27s16);
*q14s16 = vcombine_s16(d21s16, d29s16);
*q15s16 = vcombine_s16(d23s16, d31s16);
q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
vreinterpretq_s32_s16(*q10s16));
q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
vreinterpretq_s32_s16(*q11s16));
q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
vreinterpretq_s32_s16(*q14s16));
q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
vreinterpretq_s32_s16(*q15s16));
q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]), // q8
vreinterpretq_s16_s32(q1x2s32.val[0])); // q9
q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]), // q10
vreinterpretq_s16_s32(q1x2s32.val[1])); // q11
q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]), // q12
vreinterpretq_s16_s32(q3x2s32.val[0])); // q13
q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]), // q14
vreinterpretq_s16_s32(q3x2s32.val[1])); // q15
*q8s16 = q0x2s16.val[0];
*q9s16 = q0x2s16.val[1];
*q10s16 = q1x2s16.val[0];
*q11s16 = q1x2s16.val[1];
*q12s16 = q2x2s16.val[0];
*q13s16 = q2x2s16.val[1];
*q14s16 = q3x2s16.val[0];
*q15s16 = q3x2s16.val[1];
return;
}
int16x8x2_t test_vtrnq_s16(int16x8_t a, int16x8_t b) {
// CHECK-LABEL: test_vtrnq_s16
return vtrnq_s16(a, b);
// CHECK: trn1 {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
// CHECK: trn2 {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
}
void vp8_short_fdct8x4_neon(
int16_t *input,
int16_t *output,
int pitch) {
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
uint16x4_t d28u16, d29u16;
uint16x8_t q14u16;
int16x8_t q0s16, q1s16, q2s16, q3s16;
int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
int16x8x2_t v2tmp0, v2tmp1;
int32x4x2_t v2tmp2, v2tmp3;
d16s16 = vdup_n_s16(5352);
d17s16 = vdup_n_s16(2217);
q9s32 = vdupq_n_s32(14500);
q10s32 = vdupq_n_s32(7500);
// Part one
pitch >>= 1;
q0s16 = vld1q_s16(input);
input += pitch;
q1s16 = vld1q_s16(input);
input += pitch;
q2s16 = vld1q_s16(input);
input += pitch;
q3s16 = vld1q_s16(input);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q11s16 = vshlq_n_s16(q11s16, 3);
q12s16 = vshlq_n_s16(q12s16, 3);
q13s16 = vshlq_n_s16(q13s16, 3);
q14s16 = vshlq_n_s16(q14s16, 3);
q0s16 = vaddq_s16(q11s16, q12s16);
q2s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d2s16 = vshrn_n_s32(q9s32, 12);
d6s16 = vshrn_n_s32(q10s32, 12);
d3s16 = vshrn_n_s32(q11s32, 12);
d7s16 = vshrn_n_s32(q12s32, 12);
q1s16 = vcombine_s16(d2s16, d3s16);
q3s16 = vcombine_s16(d6s16, d7s16);
// Part two
q9s32 = vdupq_n_s32(12000);
q10s32 = vdupq_n_s32(51000);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q15s16 = vdupq_n_s16(7);
q11s16 = vaddq_s16(q11s16, q15s16);
q0s16 = vaddq_s16(q11s16, q12s16);
q1s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d0s16 = vget_low_s16(q0s16);
d1s16 = vget_high_s16(q0s16);
d2s16 = vget_low_s16(q1s16);
d3s16 = vget_high_s16(q1s16);
d0s16 = vshr_n_s16(d0s16, 4);
d4s16 = vshr_n_s16(d1s16, 4);
d2s16 = vshr_n_s16(d2s16, 4);
d6s16 = vshr_n_s16(d3s16, 4);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d1s16 = vshrn_n_s32(q9s32, 16);
d3s16 = vshrn_n_s32(q10s32, 16);
d5s16 = vshrn_n_s32(q11s32, 16);
d7s16 = vshrn_n_s32(q12s32, 16);
qEmptys16 = vdupq_n_s16(0);
q14u16 = vceqq_s16(q14s16, qEmptys16);
q14u16 = vmvnq_u16(q14u16);
d28u16 = vget_low_u16(q14u16);
d29u16 = vget_high_u16(q14u16);
d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
q0s16 = vcombine_s16(d0s16, d1s16);
q1s16 = vcombine_s16(d2s16, d3s16);
q2s16 = vcombine_s16(d4s16, d5s16);
q3s16 = vcombine_s16(d6s16, d7s16);
vst1q_s16(output, q0s16);
vst1q_s16(output + 8, q1s16);
vst1q_s16(output + 16, q2s16);
vst1q_s16(output + 24, q3s16);
return;
}
// TODO(johannkoenig): Make a transpose library and dedup with idct. Consider
// reversing transpose order which may make it easier for the compiler to
// reconcile the vtrn.64 moves.
static void transpose8x8(int16x8_t *a0, int16x8_t *a1,
int16x8_t *a2, int16x8_t *a3,
int16x8_t *a4, int16x8_t *a5,
int16x8_t *a6, int16x8_t *a7) {
// Swap 64 bit elements. Goes from:
// a0: 00 01 02 03 04 05 06 07
// a1: 08 09 10 11 12 13 14 15
// a2: 16 17 18 19 20 21 22 23
// a3: 24 25 26 27 28 29 30 31
// a4: 32 33 34 35 36 37 38 39
// a5: 40 41 42 43 44 45 46 47
// a6: 48 49 50 51 52 53 54 55
// a7: 56 57 58 59 60 61 62 63
// to:
// a04_lo: 00 01 02 03 32 33 34 35
// a15_lo: 08 09 10 11 40 41 42 43
// a26_lo: 16 17 18 19 48 49 50 51
// a37_lo: 24 25 26 27 56 57 58 59
// a04_hi: 04 05 06 07 36 37 38 39
// a15_hi: 12 13 14 15 44 45 46 47
// a26_hi: 20 21 22 23 52 53 54 55
// a37_hi: 28 29 30 31 60 61 62 63
const int16x8_t a04_lo = vcombine_s16(vget_low_s16(*a0), vget_low_s16(*a4));
const int16x8_t a15_lo = vcombine_s16(vget_low_s16(*a1), vget_low_s16(*a5));
const int16x8_t a26_lo = vcombine_s16(vget_low_s16(*a2), vget_low_s16(*a6));
const int16x8_t a37_lo = vcombine_s16(vget_low_s16(*a3), vget_low_s16(*a7));
const int16x8_t a04_hi = vcombine_s16(vget_high_s16(*a0), vget_high_s16(*a4));
const int16x8_t a15_hi = vcombine_s16(vget_high_s16(*a1), vget_high_s16(*a5));
const int16x8_t a26_hi = vcombine_s16(vget_high_s16(*a2), vget_high_s16(*a6));
const int16x8_t a37_hi = vcombine_s16(vget_high_s16(*a3), vget_high_s16(*a7));
// Swap 32 bit elements resulting in:
// a0246_lo:
// 00 01 16 17 32 33 48 49
// 02 03 18 19 34 35 50 51
// a1357_lo:
// 08 09 24 25 40 41 56 57
// 10 11 26 27 42 43 58 59
// a0246_hi:
// 04 05 20 21 36 37 52 53
// 06 07 22 23 38 39 54 55
// a1657_hi:
// 12 13 28 29 44 45 60 61
// 14 15 30 31 46 47 62 63
const int32x4x2_t a0246_lo = vtrnq_s32(vreinterpretq_s32_s16(a04_lo),
vreinterpretq_s32_s16(a26_lo));
const int32x4x2_t a1357_lo = vtrnq_s32(vreinterpretq_s32_s16(a15_lo),
vreinterpretq_s32_s16(a37_lo));
const int32x4x2_t a0246_hi = vtrnq_s32(vreinterpretq_s32_s16(a04_hi),
vreinterpretq_s32_s16(a26_hi));
const int32x4x2_t a1357_hi = vtrnq_s32(vreinterpretq_s32_s16(a15_hi),
vreinterpretq_s32_s16(a37_hi));
// Swap 16 bit elements resulting in:
// b0:
// 00 08 16 24 32 40 48 56
// 01 09 17 25 33 41 49 57
// b1:
// 02 10 18 26 34 42 50 58
// 03 11 19 27 35 43 51 59
// b2:
// 04 12 20 28 36 44 52 60
// 05 13 21 29 37 45 53 61
// b3:
// 06 14 22 30 38 46 54 62
// 07 15 23 31 39 47 55 63
const int16x8x2_t b0 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[0]),
vreinterpretq_s16_s32(a1357_lo.val[0]));
const int16x8x2_t b1 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[1]),
vreinterpretq_s16_s32(a1357_lo.val[1]));
const int16x8x2_t b2 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[0]),
vreinterpretq_s16_s32(a1357_hi.val[0]));
const int16x8x2_t b3 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[1]),
vreinterpretq_s16_s32(a1357_hi.val[1]));
*a0 = b0.val[0];
*a1 = b0.val[1];
*a2 = b1.val[0];
*a3 = b1.val[1];
*a4 = b2.val[0];
*a5 = b2.val[1];
*a6 = b3.val[0];
*a7 = b3.val[1];
}
void idct_dequant_full_2x_neon(
int16_t *q,
int16_t *dq,
unsigned char *dst,
int stride) {
unsigned char *dst0, *dst1;
int32x2_t d28, d29, d30, d31;
int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11;
int16x8_t qEmpty = vdupq_n_s16(0);
int32x4x2_t q2tmp0, q2tmp1;
int16x8x2_t q2tmp2, q2tmp3;
int16x4_t dLow0, dLow1, dHigh0, dHigh1;
d28 = d29 = d30 = d31 = vdup_n_s32(0);
// load dq
q0 = vld1q_s16(dq);
dq += 8;
q1 = vld1q_s16(dq);
// load q
q2 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q3 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q4 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q5 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
// load src from dst
dst0 = dst;
dst1 = dst + 4;
d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0);
dst0 += stride;
d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1);
dst1 += stride;
d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0);
dst0 += stride;
d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1);
dst1 += stride;
d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0);
dst0 += stride;
d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1);
dst1 += stride;
d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0);
d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1);
q2 = vmulq_s16(q2, q0);
q3 = vmulq_s16(q3, q1);
q4 = vmulq_s16(q4, q0);
q5 = vmulq_s16(q5, q1);
// vswp
dLow0 = vget_low_s16(q2);
dHigh0 = vget_high_s16(q2);
dLow1 = vget_low_s16(q4);
dHigh1 = vget_high_s16(q4);
q2 = vcombine_s16(dLow0, dLow1);
q4 = vcombine_s16(dHigh0, dHigh1);
dLow0 = vget_low_s16(q3);
dHigh0 = vget_high_s16(q3);
dLow1 = vget_low_s16(q5);
dHigh1 = vget_high_s16(q5);
q3 = vcombine_s16(dLow0, dLow1);
q5 = vcombine_s16(dHigh0, dHigh1);
q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2);
q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2);
q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1);
q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1);
q10 = vqaddq_s16(q2, q3);
q11 = vqsubq_s16(q2, q3);
q8 = vshrq_n_s16(q8, 1);
q9 = vshrq_n_s16(q9, 1);
q4 = vqaddq_s16(q4, q8);
q5 = vqaddq_s16(q5, q9);
q2 = vqsubq_s16(q6, q5);
q3 = vqaddq_s16(q7, q4);
q4 = vqaddq_s16(q10, q3);
q5 = vqaddq_s16(q11, q2);
q6 = vqsubq_s16(q11, q2);
q7 = vqsubq_s16(q10, q3);
q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
vreinterpretq_s16_s32(q2tmp1.val[0]));
q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
vreinterpretq_s16_s32(q2tmp1.val[1]));
// loop 2
q8 = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2);
q9 = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2);
q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1);
q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1);
q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]);
q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]);
q10 = vshrq_n_s16(q10, 1);
q11 = vshrq_n_s16(q11, 1);
q10 = vqaddq_s16(q2tmp2.val[1], q10);
q11 = vqaddq_s16(q2tmp3.val[1], q11);
q8 = vqsubq_s16(q8, q11);
q9 = vqaddq_s16(q9, q10);
q4 = vqaddq_s16(q2, q9);
q5 = vqaddq_s16(q3, q8);
q6 = vqsubq_s16(q3, q8);
q7 = vqsubq_s16(q2, q9);
q4 = vrshrq_n_s16(q4, 3);
q5 = vrshrq_n_s16(q5, 3);
q6 = vrshrq_n_s16(q6, 3);
q7 = vrshrq_n_s16(q7, 3);
q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
vreinterpretq_s16_s32(q2tmp1.val[0]));
q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
vreinterpretq_s16_s32(q2tmp1.val[1]));
q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]),
vreinterpret_u8_s32(d28)));
q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]),
vreinterpret_u8_s32(d29)));
q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]),
vreinterpret_u8_s32(d30)));
q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]),
vreinterpret_u8_s32(d31)));
d28 = vreinterpret_s32_u8(vqmovun_s16(q4));
d29 = vreinterpret_s32_u8(vqmovun_s16(q5));
d30 = vreinterpret_s32_u8(vqmovun_s16(q6));
d31 = vreinterpret_s32_u8(vqmovun_s16(q7));
dst0 = dst;
dst1 = dst + 4;
vst1_lane_s32((int32_t *)dst0, d28, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d28, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d29, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d29, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d30, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d30, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d31, 0);
vst1_lane_s32((int32_t *)dst1, d31, 1);
return;
}