static INLINE void GENERATE_COSINE_CONSTANTS(int16x4_t *d0s16, int16x4_t *d1s16,
int16x4_t *d2s16) {
*d0s16 = vdup_n_s16(cospi_8_64);
*d1s16 = vdup_n_s16(cospi_16_64);
*d2s16 = vdup_n_s16(cospi_24_64);
return;
}
static INLINE void GENERATE_SINE_CONSTANTS(int16x4_t *d3s16, int16x4_t *d4s16,
int16x4_t *d5s16, int16x8_t *q3s16) {
*d3s16 = vdup_n_s16(sinpi_1_9);
*d4s16 = vdup_n_s16(sinpi_2_9);
*q3s16 = vdupq_n_s16(sinpi_3_9);
*d5s16 = vdup_n_s16(sinpi_4_9);
return;
}
int16x4_t sub_abs_to_vabd_16()
{
int16x4_t val1 = vdup_n_s16 (10);
int16x4_t val2 = vdup_n_s16 (30);
int16x4_t sres = vsub_s16(val1, val2);
int16x4_t res = vabs_s16 (sres);
return res;
}
void test_vdup_ns16 (void)
{
int16x4_t out_int16x4_t;
int16_t arg0_int16_t;
out_int16x4_t = vdup_n_s16 (arg0_int16_t);
}
/* return the sum of all elements in an array. This works by calculating 4 totals (one for each lane) and adding those at the end to get the final total */
int sum_array(int16_t *array, int size)
{
/* initialize the accumulator vector to zero */
int16x4_t acc = vdup_n_s16(0);
int32x2_t acc1;
int64x1_t acc2;
/* this implementation assumes the size of the array is a multiple of 4 */
assert((size % 4) == 0);
/* counting backwards gives better code */
for (; size != 0; size -= 4)
{
int16x4_t vec;
/* load 4 values in parallel from the array */
vec = vld1_s16(array);
/* increment the array pointer to the next element */
array += 4;
/* add the vector to the accumulator vector */
acc = vadd_s16(acc, vec);
}
/* calculate the total */
acc1 = vpaddl_s16(acc);
acc2 = vpaddl_s32(acc1);
/* return the total as an integer */
return (int)vget_lane_s64(acc2, 0);
}
static void inline ff_dct_unquantize_h263_neon(int qscale, int qadd, int nCoeffs,
int16_t *block)
{
int16x8_t q0s16, q2s16, q3s16, q8s16, q10s16, q11s16, q13s16;
int16x8_t q14s16, q15s16, qzs16;
int16x4_t d0s16, d2s16, d3s16, dzs16;
uint16x8_t q1u16, q9u16;
uint16x4_t d1u16;
dzs16 = vdup_n_s16(0);
qzs16 = vdupq_n_s16(0);
q15s16 = vdupq_n_s16(qscale << 1);
q14s16 = vdupq_n_s16(qadd);
q13s16 = vnegq_s16(q14s16);
if (nCoeffs > 4) {
for (; nCoeffs > 8; nCoeffs -= 16, block += 16) {
q0s16 = vld1q_s16(block);
q3s16 = vreinterpretq_s16_u16(vcltq_s16(q0s16, qzs16));
q8s16 = vld1q_s16(block + 8);
q1u16 = vceqq_s16(q0s16, qzs16);
q2s16 = vmulq_s16(q0s16, q15s16);
q11s16 = vreinterpretq_s16_u16(vcltq_s16(q8s16, qzs16));
q10s16 = vmulq_s16(q8s16, q15s16);
q3s16 = vbslq_s16(vreinterpretq_u16_s16(q3s16), q13s16, q14s16);
q11s16 = vbslq_s16(vreinterpretq_u16_s16(q11s16), q13s16, q14s16);
q2s16 = vaddq_s16(q2s16, q3s16);
q9u16 = vceqq_s16(q8s16, qzs16);
q10s16 = vaddq_s16(q10s16, q11s16);
q0s16 = vbslq_s16(q1u16, q0s16, q2s16);
q8s16 = vbslq_s16(q9u16, q8s16, q10s16);
vst1q_s16(block, q0s16);
vst1q_s16(block + 8, q8s16);
}
}
if (nCoeffs <= 0)
return;
d0s16 = vld1_s16(block);
d3s16 = vreinterpret_s16_u16(vclt_s16(d0s16, dzs16));
d1u16 = vceq_s16(d0s16, dzs16);
d2s16 = vmul_s16(d0s16, vget_high_s16(q15s16));
d3s16 = vbsl_s16(vreinterpret_u16_s16(d3s16),
vget_high_s16(q13s16), vget_high_s16(q14s16));
d2s16 = vadd_s16(d2s16, d3s16);
d0s16 = vbsl_s16(d1u16, d0s16, d2s16);
vst1_s16(block, d0s16);
}
inline void ClampBufferToS16(s16 *out, const s32 *in, size_t size, s8 volShift) {
#ifdef _M_SSE
// Size will always be 16-byte aligned as the hwBlockSize is.
while (size >= 8) {
__m128i in1 = _mm_loadu_si128((__m128i *)in);
__m128i in2 = _mm_loadu_si128((__m128i *)(in + 4));
__m128i packed = _mm_packs_epi32(in1, in2);
if (useShift) {
packed = _mm_srai_epi16(packed, volShift);
}
_mm_storeu_si128((__m128i *)out, packed);
out += 8;
in += 8;
size -= 8;
}
#elif PPSSPP_ARCH(ARM_NEON)
int16x4_t signedVolShift = vdup_n_s16 (-volShift); // Can only dynamic-shift right, but by a signed integer
while (size >= 8) {
int32x4_t in1 = vld1q_s32(in);
int32x4_t in2 = vld1q_s32(in + 4);
int16x4_t packed1 = vqmovn_s32(in1);
int16x4_t packed2 = vqmovn_s32(in2);
if (useShift) {
packed1 = vshl_s16(packed1, signedVolShift);
packed2 = vshl_s16(packed2, signedVolShift);
}
vst1_s16(out, packed1);
vst1_s16(out + 4, packed2);
out += 8;
in += 8;
size -= 8;
}
#endif
// This does the remainder if SIMD was used, otherwise it does it all.
for (size_t i = 0; i < size; i++) {
out[i] = clamp_s16(useShift ? (in[i] >> volShift) : in[i]);
}
}
// coeff: 16 bits, dynamic range [-32640, 32640].
// length: value range {16, 64, 256, 1024}.
int aom_satd_neon(const int16_t *coeff, int length) {
const int16x4_t zero = vdup_n_s16(0);
int32x4_t accum = vdupq_n_s32(0);
do {
const int16x8_t src0 = vld1q_s16(coeff);
const int16x8_t src8 = vld1q_s16(coeff + 8);
accum = vabal_s16(accum, vget_low_s16(src0), zero);
accum = vabal_s16(accum, vget_high_s16(src0), zero);
accum = vabal_s16(accum, vget_low_s16(src8), zero);
accum = vabal_s16(accum, vget_high_s16(src8), zero);
length -= 16;
coeff += 16;
} while (length != 0);
{
// satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
const int64x2_t s0 = vpaddlq_s32(accum); // cascading summation of 'accum'.
const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
vreinterpret_s32_s64(vget_high_s64(s0)));
const int satd = vget_lane_s32(s1, 0);
return satd;
}
}
static INLINE void IDCT8x8_1D(
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 d0s16, d1s16, d2s16, d3s16;
int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
d0s16 = vdup_n_s16(cospi_28_64);
d1s16 = vdup_n_s16(cospi_4_64);
d2s16 = vdup_n_s16(cospi_12_64);
d3s16 = vdup_n_s16(cospi_20_64);
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);
q2s32 = vmull_s16(d18s16, d0s16);
q3s32 = vmull_s16(d19s16, d0s16);
q5s32 = vmull_s16(d26s16, d2s16);
q6s32 = vmull_s16(d27s16, d2s16);
q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
d8s16 = vqrshrn_n_s32(q2s32, 14);
d9s16 = vqrshrn_n_s32(q3s32, 14);
d10s16 = vqrshrn_n_s32(q5s32, 14);
d11s16 = vqrshrn_n_s32(q6s32, 14);
q4s16 = vcombine_s16(d8s16, d9s16);
q5s16 = vcombine_s16(d10s16, d11s16);
q2s32 = vmull_s16(d18s16, d1s16);
q3s32 = vmull_s16(d19s16, d1s16);
q9s32 = vmull_s16(d26s16, d3s16);
q13s32 = vmull_s16(d27s16, d3s16);
q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
d14s16 = vqrshrn_n_s32(q2s32, 14);
d15s16 = vqrshrn_n_s32(q3s32, 14);
d12s16 = vqrshrn_n_s32(q9s32, 14);
d13s16 = vqrshrn_n_s32(q13s32, 14);
q6s16 = vcombine_s16(d12s16, d13s16);
q7s16 = vcombine_s16(d14s16, d15s16);
d0s16 = vdup_n_s16(cospi_16_64);
q2s32 = vmull_s16(d16s16, d0s16);
q3s32 = vmull_s16(d17s16, d0s16);
q13s32 = vmull_s16(d16s16, d0s16);
q15s32 = vmull_s16(d17s16, d0s16);
q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
d0s16 = vdup_n_s16(cospi_24_64);
d1s16 = vdup_n_s16(cospi_8_64);
d18s16 = vqrshrn_n_s32(q2s32, 14);
d19s16 = vqrshrn_n_s32(q3s32, 14);
d22s16 = vqrshrn_n_s32(q13s32, 14);
d23s16 = vqrshrn_n_s32(q15s32, 14);
*q9s16 = vcombine_s16(d18s16, d19s16);
*q11s16 = vcombine_s16(d22s16, d23s16);
q2s32 = vmull_s16(d20s16, d0s16);
q3s32 = vmull_s16(d21s16, d0s16);
q8s32 = vmull_s16(d20s16, d1s16);
q12s32 = vmull_s16(d21s16, d1s16);
q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
d26s16 = vqrshrn_n_s32(q2s32, 14);
d27s16 = vqrshrn_n_s32(q3s32, 14);
d30s16 = vqrshrn_n_s32(q8s32, 14);
d31s16 = vqrshrn_n_s32(q12s32, 14);
*q13s16 = vcombine_s16(d26s16, d27s16);
*q15s16 = vcombine_s16(d30s16, d31s16);
q0s16 = vaddq_s16(*q9s16, *q15s16);
q1s16 = vaddq_s16(*q11s16, *q13s16);
q2s16 = vsubq_s16(*q11s16, *q13s16);
q3s16 = vsubq_s16(*q9s16, *q15s16);
*q13s16 = vsubq_s16(q4s16, q5s16);
q4s16 = vaddq_s16(q4s16, q5s16);
*q14s16 = vsubq_s16(q7s16, q6s16);
q7s16 = vaddq_s16(q7s16, q6s16);
d26s16 = vget_low_s16(*q13s16);
d27s16 = vget_high_s16(*q13s16);
d28s16 = vget_low_s16(*q14s16);
d29s16 = vget_high_s16(*q14s16);
d16s16 = vdup_n_s16(cospi_16_64);
q9s32 = vmull_s16(d28s16, d16s16);
q10s32 = vmull_s16(d29s16, d16s16);
q11s32 = vmull_s16(d28s16, d16s16);
q12s32 = vmull_s16(d29s16, d16s16);
q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
d10s16 = vqrshrn_n_s32(q9s32, 14);
d11s16 = vqrshrn_n_s32(q10s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q12s32, 14);
q5s16 = vcombine_s16(d10s16, d11s16);
q6s16 = vcombine_s16(d12s16, d13s16);
*q8s16 = vaddq_s16(q0s16, q7s16);
*q9s16 = vaddq_s16(q1s16, q6s16);
*q10s16 = vaddq_s16(q2s16, q5s16);
*q11s16 = vaddq_s16(q3s16, q4s16);
*q12s16 = vsubq_s16(q3s16, q4s16);
*q13s16 = vsubq_s16(q2s16, q5s16);
*q14s16 = vsubq_s16(q1s16, q6s16);
*q15s16 = vsubq_s16(q0s16, q7s16);
return;
}
void vpx_idct8x8_12_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
int16x4_t d26s16, d27s16, d28s16, d29s16;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 16);
q11s16 = vld1q_s16(input + 24);
q12s16 = vld1q_s16(input + 32);
q13s16 = vld1q_s16(input + 40);
q14s16 = vld1q_s16(input + 48);
q15s16 = vld1q_s16(input + 56);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
// First transform rows
// stage 1
q0s16 = vdupq_n_s16(cospi_28_64 * 2);
q1s16 = vdupq_n_s16(cospi_4_64 * 2);
q4s16 = vqrdmulhq_s16(q9s16, q0s16);
q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
q7s16 = vqrdmulhq_s16(q9s16, q1s16);
q1s16 = vdupq_n_s16(cospi_12_64 * 2);
q5s16 = vqrdmulhq_s16(q11s16, q0s16);
q0s16 = vdupq_n_s16(cospi_16_64 * 2);
q6s16 = vqrdmulhq_s16(q11s16, q1s16);
// stage 2 & stage 3 - even half
q1s16 = vdupq_n_s16(cospi_24_64 * 2);
q9s16 = vqrdmulhq_s16(q8s16, q0s16);
q0s16 = vdupq_n_s16(cospi_8_64 * 2);
q13s16 = vqrdmulhq_s16(q10s16, q1s16);
q15s16 = vqrdmulhq_s16(q10s16, q0s16);
// stage 3 -odd half
q0s16 = vaddq_s16(q9s16, q15s16);
q1s16 = vaddq_s16(q9s16, q13s16);
q2s16 = vsubq_s16(q9s16, q13s16);
q3s16 = vsubq_s16(q9s16, q15s16);
// stage 2 - odd half
q13s16 = vsubq_s16(q4s16, q5s16);
q4s16 = vaddq_s16(q4s16, q5s16);
q14s16 = vsubq_s16(q7s16, q6s16);
q7s16 = vaddq_s16(q7s16, q6s16);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
d16s16 = vdup_n_s16(cospi_16_64);
q9s32 = vmull_s16(d28s16, d16s16);
q10s32 = vmull_s16(d29s16, d16s16);
q11s32 = vmull_s16(d28s16, d16s16);
q12s32 = vmull_s16(d29s16, d16s16);
q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
d10s16 = vqrshrn_n_s32(q9s32, 14);
d11s16 = vqrshrn_n_s32(q10s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q12s32, 14);
q5s16 = vcombine_s16(d10s16, d11s16);
q6s16 = vcombine_s16(d12s16, d13s16);
// stage 4
q8s16 = vaddq_s16(q0s16, q7s16);
q9s16 = vaddq_s16(q1s16, q6s16);
q10s16 = vaddq_s16(q2s16, q5s16);
q11s16 = vaddq_s16(q3s16, q4s16);
q12s16 = vsubq_s16(q3s16, q4s16);
q13s16 = vsubq_s16(q2s16, q5s16);
q14s16 = vsubq_s16(q1s16, q6s16);
q15s16 = vsubq_s16(q0s16, q7s16);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
d1 = d2 = dest;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
return;
}
void ne10_img_hresize_4channels_linear_neon (const unsigned char** src, int** dst, int count,
const int* xofs, const short* alpha,
int swidth, int dwidth, int cn, int xmin, int xmax)
{
int dx, k;
int dx0 = 0;
int16x4x2_t alpha_vec;
uint8x8_t dS0_vec, dS1_vec;
int16x8_t qS0_vec, qS1_vec;
int16x4_t dS0_0123, dS0_4567, dS1_0123, dS1_4567;
int32x4_t qT0_vec, qT1_vec;
int16x4_t dCoeff;
dCoeff = vdup_n_s16 (INTER_RESIZE_COEF_SCALE);
for (k = 0; k <= count - 2; k++)
{
const unsigned char *S0 = src[k], *S1 = src[k + 1];
int *D0 = dst[k], *D1 = dst[k + 1];
for (dx = dx0; dx < xmax; dx += 4)
{
int sx = xofs[dx];
alpha_vec = vld2_s16 (&alpha[dx * 2]);
dS0_vec = vld1_u8 (&S0[sx]);
dS1_vec = vld1_u8 (&S1[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
qS1_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS1_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS0_4567 = vget_high_s16 (qS0_vec);
dS1_0123 = vget_low_s16 (qS1_vec);
dS1_4567 = vget_high_s16 (qS1_vec);
qT0_vec = vmull_s16 (dS0_0123, alpha_vec.val[0]);
qT1_vec = vmull_s16 (dS1_0123, alpha_vec.val[0]);
qT0_vec = vmlal_s16 (qT0_vec, dS0_4567, alpha_vec.val[1]);
qT1_vec = vmlal_s16 (qT1_vec, dS1_4567, alpha_vec.val[1]);
vst1q_s32 (&D0[dx], qT0_vec);
vst1q_s32 (&D1[dx], qT1_vec);
}
for (; dx < dwidth; dx += 4)
{
int sx = xofs[dx];
dS0_vec = vld1_u8 (&S0[sx]);
dS1_vec = vld1_u8 (&S1[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
qS1_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS1_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS1_0123 = vget_low_s16 (qS1_vec);
qT0_vec = vmull_s16 (dS0_0123, dCoeff);
qT1_vec = vmull_s16 (dS1_0123, dCoeff);
vst1q_s32 (&D0[dx], qT0_vec);
vst1q_s32 (&D1[dx], qT1_vec);
}
}
for (; k < count; k++)
{
const unsigned char *S = src[k];
int *D = dst[k];
for (dx = 0; dx < xmax; dx += 4)
{
int sx = xofs[dx];
alpha_vec = vld2_s16 (&alpha[dx * 2]);
dS0_vec = vld1_u8 (&S[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
dS0_4567 = vget_high_s16 (qS0_vec);
qT0_vec = vmull_s16 (dS0_0123, alpha_vec.val[0]);
qT0_vec = vmlal_s16 (qT0_vec, dS0_4567, alpha_vec.val[1]);
vst1q_s32 (&D[dx], qT0_vec);
}
for (; dx < dwidth; dx += 4)
{
int sx = xofs[dx];
dS0_vec = vld1_u8 (&S[sx]);
qS0_vec = vreinterpretq_s16_u16 (vmovl_u8 (dS0_vec));
dS0_0123 = vget_low_s16 (qS0_vec);
qT0_vec = vmull_s16 (dS0_0123, dCoeff);
vst1q_s32 (&D[dx], qT0_vec);
}
}
}
int16x4_t test_vdup_n_s16(int16_t v1) {
// CHECK: test_vdup_n_s16
return vdup_n_s16(v1);
// CHECK: dup {{v[0-9]+}}.4h, {{w[0-9]+}}
}
void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
uint8x8_t d26u8, d27u8;
uint32x2_t d26u32, d27u32;
uint16x8_t q8u16, q9u16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
int16x8_t q8s16, q9s16, q13s16, q14s16;
int32x4_t q1s32, q13s32, q14s32, q15s32;
int16x4x2_t d0x2s16, d1x2s16;
int32x4x2_t q0x2s32;
uint8_t *d;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_low_s16(q9s16);
d19s16 = vget_high_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
d20s16 = vdup_n_s16((int16_t)cospi_8_64);
d21s16 = vdup_n_s16((int16_t)cospi_16_64);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d22s16 = vdup_n_s16((int16_t)cospi_24_64);
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_high_s16(q9s16); // vswp d18 d19
d19s16 = vget_low_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
// do the transform on columns
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d = dest;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
d += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d = dest;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
return;
}
// Update the noise estimation information.
static void UpdateNoiseEstimateNeon(NoiseSuppressionFixedC* inst, int offset) {
const int16_t kExp2Const = 11819; // Q13
int16_t* ptr_noiseEstLogQuantile = NULL;
int16_t* ptr_noiseEstQuantile = NULL;
int16x4_t kExp2Const16x4 = vdup_n_s16(kExp2Const);
int32x4_t twentyOne32x4 = vdupq_n_s32(21);
int32x4_t constA32x4 = vdupq_n_s32(0x1fffff);
int32x4_t constB32x4 = vdupq_n_s32(0x200000);
int16_t tmp16 = WebRtcSpl_MaxValueW16(inst->noiseEstLogQuantile + offset,
inst->magnLen);
// Guarantee a Q-domain as high as possible and still fit in int16
inst->qNoise = 14 - (int) WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(kExp2Const,
tmp16,
21);
int32x4_t qNoise32x4 = vdupq_n_s32(inst->qNoise);
for (ptr_noiseEstLogQuantile = &inst->noiseEstLogQuantile[offset],
ptr_noiseEstQuantile = &inst->noiseEstQuantile[0];
ptr_noiseEstQuantile < &inst->noiseEstQuantile[inst->magnLen - 3];
ptr_noiseEstQuantile += 4, ptr_noiseEstLogQuantile += 4) {
// tmp32no2 = kExp2Const * inst->noiseEstLogQuantile[offset + i];
int16x4_t v16x4 = vld1_s16(ptr_noiseEstLogQuantile);
int32x4_t v32x4B = vmull_s16(v16x4, kExp2Const16x4);
// tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
int32x4_t v32x4A = vandq_s32(v32x4B, constA32x4);
v32x4A = vorrq_s32(v32x4A, constB32x4);
// tmp16 = (int16_t)(tmp32no2 >> 21);
v32x4B = vshrq_n_s32(v32x4B, 21);
// tmp16 -= 21;// shift 21 to get result in Q0
v32x4B = vsubq_s32(v32x4B, twentyOne32x4);
// tmp16 += (int16_t) inst->qNoise;
// shift to get result in Q(qNoise)
v32x4B = vaddq_s32(v32x4B, qNoise32x4);
// if (tmp16 < 0) {
// tmp32no1 >>= -tmp16;
// } else {
// tmp32no1 <<= tmp16;
// }
v32x4B = vshlq_s32(v32x4A, v32x4B);
// tmp16 = WebRtcSpl_SatW32ToW16(tmp32no1);
v16x4 = vqmovn_s32(v32x4B);
//inst->noiseEstQuantile[i] = tmp16;
vst1_s16(ptr_noiseEstQuantile, v16x4);
}
// Last iteration:
// inst->quantile[i]=exp(inst->lquantile[offset+i]);
// in Q21
int32_t tmp32no2 = kExp2Const * *ptr_noiseEstLogQuantile;
int32_t tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
tmp16 = (int16_t)(tmp32no2 >> 21);
tmp16 -= 21;// shift 21 to get result in Q0
tmp16 += (int16_t) inst->qNoise; //shift to get result in Q(qNoise)
if (tmp16 < 0) {
tmp32no1 >>= -tmp16;
} else {
frac = (int16_t)((((uint32_t)magn[i] << zeros)
& 0x7FFFFFFF) >> 23);
assert(frac < 256);
// log2(magn(i))
log2 = (int16_t)(((31 - zeros) << 8)
+ WebRtcNsx_kLogTableFrac[frac]);
// log2(magn(i))*log(2)
lmagn[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(log2, log2_const, 15);
// + log(2^stages)
lmagn[i] += logval;
} else {
lmagn[i] = logval;
}
}
int16x4_t Q3_16x4 = vdup_n_s16(3);
int16x8_t WIDTHQ8_16x8 = vdupq_n_s16(WIDTH_Q8);
int16x8_t WIDTHFACTOR_16x8 = vdupq_n_s16(width_factor);
int16_t factor = FACTOR_Q7;
if (inst->blockIndex < END_STARTUP_LONG)
factor = FACTOR_Q7_STARTUP;
// Loop over simultaneous estimates
for (s = 0; s < SIMULT; s++) {
offset = s * inst->magnLen;
// Get counter values from state
counter = inst->noiseEstCounter[s];
assert(counter < 201);
countDiv = WebRtcNsx_kCounterDiv[counter];
void silk_biquad_alt_stride2_neon(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
)
{
/* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
opus_int k = 0;
const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 );
const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 );
int16x4_t in_s16x4 = vdup_n_s16( 0 );
int16x4_t out_s16x4;
int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2;
int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4;
int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2;
#ifdef OPUS_CHECK_ASM
opus_int32 S_c[ 4 ];
VARDECL( opus_int16, out_c );
SAVE_STACK;
ALLOC( out_c, 2 * len, opus_int16 );
silk_memcpy( &S_c, S, sizeof( S_c ) );
silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len );
#endif
/* Negate A_Q28 values and split in two parts */
A_Q28_s32x2 = vld1_s32( A_Q28 );
A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 );
A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */
A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */
A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */
A_U_s32x2 = vshl_n_s32( A_U_s32x2, 16 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 16 (Clip two leading bits to conform to C function.) */
A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */
B_Q28_s32x2 = vld1_s32( B_Q28 );
t_s32x2 = vld1_s32( B_Q28 + 1 );
t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 );
t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 );
t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 );
A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */
A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */
B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */
S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */
S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */
S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] );
for( ; k < len - 1; k += 2 ) {
int32x4_t in_s32x4[ 2 ], t_s32x4;
int32x2_t out32_Q14_s32x2[ 2 ];
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */
in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */
in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */
t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */
in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */
in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_low_s32 ( t_s32x4 ), in_s32x4[ 0 ], &S_s32x4, &out32_Q14_s32x2[ 0 ] );
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_high_s32( t_s32x4 ), in_s32x4[ 1 ], &S_s32x4, &out32_Q14_s32x2[ 1 ] );
/* Scale back to Q0 and saturate */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */
out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */
out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ) */
vst1_s16( &out[ 2 * k ], out_s16x4 ); /* out[ 2 * k + {0,1,2,3} ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ); */
}
/* Process leftover. */
if( k < len ) {
int32x4_t in_s32x4;
int32x2_t out32_Q14_s32x2;
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */
t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */
in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 );
/* Scale back to Q0 and saturate */
out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */
out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,0,1} + (1<<14) - 1, 14 ) ) */
vst1_lane_s16( &out[ 2 * k + 0 ], out_s16x4, 0 ); /* out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_0 + (1<<14) - 1, 14 ) ); */
vst1_lane_s16( &out[ 2 * k + 1 ], out_s16x4, 1 ); /* out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_1 + (1<<14) - 1, 14 ) ); */
}
vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */
vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */
vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */
vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */
#ifdef OPUS_CHECK_ASM
silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) );
silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) );
RESTORE_STACK;
#endif
}
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;
}
inline int16x4_t vdup_n(const s16 & val) { return vdup_n_s16(val); }
void vp8_short_fdct4x4_neon(
int16_t *input,
int16_t *output,
int pitch) {
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
int16x4_t d16s16, d17s16, d26s16, dEmptys16;
uint16x4_t d4u16;
int16x8_t q0s16, q1s16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
int16x4x2_t v2tmp0, v2tmp1;
int32x2x2_t v2tmp2, v2tmp3;
d16s16 = vdup_n_s16(5352);
d17s16 = vdup_n_s16(2217);
q9s32 = vdupq_n_s32(14500);
q10s32 = vdupq_n_s32(7500);
q11s32 = vdupq_n_s32(12000);
q12s32 = vdupq_n_s32(51000);
// Part one
pitch >>= 1;
d0s16 = vld1_s16(input);
input += pitch;
d1s16 = vld1_s16(input);
input += pitch;
d2s16 = vld1_s16(input);
input += pitch;
d3s16 = vld1_s16(input);
v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
vreinterpret_s32_s16(d2s16));
v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
vreinterpret_s32_s16(d3s16));
v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]), // d0
vreinterpret_s16_s32(v2tmp3.val[0])); // d1
v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]), // d2
vreinterpret_s16_s32(v2tmp3.val[1])); // d3
d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
d4s16 = vshl_n_s16(d4s16, 3);
d5s16 = vshl_n_s16(d5s16, 3);
d6s16 = vshl_n_s16(d6s16, 3);
d7s16 = vshl_n_s16(d7s16, 3);
d0s16 = vadd_s16(d4s16, d5s16);
d2s16 = vsub_s16(d4s16, d5s16);
q9s32 = vmlal_s16(q9s32, d7s16, d16s16);
q10s32 = vmlal_s16(q10s32, d7s16, d17s16);
q9s32 = vmlal_s16(q9s32, d6s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d6s16, d16s16);
d1s16 = vshrn_n_s32(q9s32, 12);
d3s16 = vshrn_n_s32(q10s32, 12);
// Part two
v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
vreinterpret_s32_s16(d2s16));
v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
vreinterpret_s32_s16(d3s16));
v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]), // d0
vreinterpret_s16_s32(v2tmp3.val[0])); // d1
v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]), // d2
vreinterpret_s16_s32(v2tmp3.val[1])); // d3
d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
d26s16 = vdup_n_s16(7);
d4s16 = vadd_s16(d4s16, d26s16);
d0s16 = vadd_s16(d4s16, d5s16);
d2s16 = vsub_s16(d4s16, d5s16);
q11s32 = vmlal_s16(q11s32, d7s16, d16s16);
q12s32 = vmlal_s16(q12s32, d7s16, d17s16);
dEmptys16 = vdup_n_s16(0);
d4u16 = vceq_s16(d7s16, dEmptys16);
d0s16 = vshr_n_s16(d0s16, 4);
d2s16 = vshr_n_s16(d2s16, 4);
q11s32 = vmlal_s16(q11s32, d6s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d6s16, d16s16);
d4u16 = vmvn_u16(d4u16);
d1s16 = vshrn_n_s32(q11s32, 16);
d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d4u16));
d3s16 = vshrn_n_s32(q12s32, 16);
q0s16 = vcombine_s16(d0s16, d1s16);
q1s16 = vcombine_s16(d2s16, d3s16);
vst1q_s16(output, q0s16);
vst1q_s16(output + 8, q1s16);
return;
}
