static uint64_t SharpYUVUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len) {
int i;
const int16x8_t zero = vdupq_n_s16(0);
const int16x8_t max = vdupq_n_s16(MAX_Y);
uint64x2_t sum = vdupq_n_u64(0);
uint64_t diff;
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
const int16x8_t D = vsubq_s16(A, B); // diff_y
const int16x8_t F = vaddq_s16(C, D); // new_y
const uint16x8_t H =
vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
const int16x8_t I = vabsq_s16(D); // abs(diff_y)
vst1q_u16(dst + i, H);
sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
}
diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
for (; i < len; ++i) {
const int diff_y = ref[i] - src[i];
const int new_y = (int)(dst[i]) + diff_y;
dst[i] = clip_y(new_y);
diff += (uint64_t)(abs(diff_y));
}
return diff;
}
/* u16x8 mv mul */
void mw_neon_mv_mul_u16x8(unsigned short * A, int Row, int T, unsigned short * B, unsigned short * C)
{
int i = 0;
int k = 0;
uint16x8_t neon_b, neon_c;
uint16x8_t neon_a0, neon_a1, neon_a2, neon_a3, neon_a4, neon_a5, neon_a6, neon_a7;
uint16x8_t neon_b0, neon_b1, neon_b2, neon_b3, neon_b4, neon_b5, neon_b6, neon_b7;
for (i = 0; i < Row; i+=8)
{
neon_c = vmovq_n_u16(0);
for (k = 0; k < T; k+=8)
{
int j = k * T + i;
neon_a0 = vld1q_u16(A + j);
j+=Row;
neon_a1 = vld1q_u16(A + j);
j+=Row;
neon_a2 = vld1q_u16(A + j);
j+=Row;
neon_a3 = vld1q_u16(A + j);
j+=Row;
neon_a4 = vld1q_u16(A + j);
j+=Row;
neon_a5 = vld1q_u16(A + j);
j+=Row;
neon_a6 = vld1q_u16(A + j);
j+=Row;
neon_a7 = vld1q_u16(A + j);
neon_b = vld1q_u16(B + k);
neon_b0 = vdupq_n_u16(vgetq_lane_u16(neon_b, 0));
neon_b1 = vdupq_n_u16(vgetq_lane_u16(neon_b, 1));
neon_b2 = vdupq_n_u16(vgetq_lane_u16(neon_b, 2));
neon_b3 = vdupq_n_u16(vgetq_lane_u16(neon_b, 3));
neon_b4 = vdupq_n_u16(vgetq_lane_u16(neon_b, 4));
neon_b5 = vdupq_n_u16(vgetq_lane_u16(neon_b, 5));
neon_b6 = vdupq_n_u16(vgetq_lane_u16(neon_b, 6));
neon_b7 = vdupq_n_u16(vgetq_lane_u16(neon_b, 7));
neon_c = vaddq_u16(vmulq_u16(neon_a0, neon_b0), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a1, neon_b1), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a2, neon_b2), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a3, neon_b3), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a4, neon_b4), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a5, neon_b5), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a6, neon_b6), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a7, neon_b7), neon_c);
}
vst1q_u16(C + i, neon_c);
}
}
void SkBlitLCD16Row_neon(SkPMColor dst[], const uint16_t src[],
SkColor color, int width, SkPMColor) {
int colA = SkColorGetA(color);
int colR = SkColorGetR(color);
int colG = SkColorGetG(color);
int colB = SkColorGetB(color);
colA = SkAlpha255To256(colA);
uint8x8_t vcolR, vcolG, vcolB;
uint16x8_t vcolA;
if (width >= 8) {
vcolA = vdupq_n_u16(colA);
vcolR = vdup_n_u8(colR);
vcolG = vdup_n_u8(colG);
vcolB = vdup_n_u8(colB);
}
while (width >= 8) {
uint8x8x4_t vdst;
uint16x8_t vmask;
uint16x8_t vmaskR, vmaskG, vmaskB;
vdst = vld4_u8((uint8_t*)dst);
vmask = vld1q_u16(src);
// Get all the color masks on 5 bits
vmaskR = vshrq_n_u16(vmask, SK_R16_SHIFT);
vmaskG = vshrq_n_u16(vshlq_n_u16(vmask, SK_R16_BITS),
SK_B16_BITS + SK_R16_BITS + 1);
vmaskB = vmask & vdupq_n_u16(SK_B16_MASK);
// Upscale to 0..32
vmaskR = vmaskR + vshrq_n_u16(vmaskR, 4);
vmaskG = vmaskG + vshrq_n_u16(vmaskG, 4);
vmaskB = vmaskB + vshrq_n_u16(vmaskB, 4);
vmaskR = vshrq_n_u16(vmaskR * vcolA, 8);
vmaskG = vshrq_n_u16(vmaskG * vcolA, 8);
vmaskB = vshrq_n_u16(vmaskB * vcolA, 8);
vdst.val[NEON_A] = vdup_n_u8(0xFF);
vdst.val[NEON_R] = SkBlend32_neon8(vcolR, vdst.val[NEON_R], vmaskR);
vdst.val[NEON_G] = SkBlend32_neon8(vcolG, vdst.val[NEON_G], vmaskG);
vdst.val[NEON_B] = SkBlend32_neon8(vcolB, vdst.val[NEON_B], vmaskB);
vst4_u8((uint8_t*)dst, vdst);
dst += 8;
src += 8;
width -= 8;
}
for (int i = 0; i < width; i++) {
dst[i] = SkBlendLCD16(colA, colR, colG, colB, dst[i], src[i]);
}
}
void WebRtcAecm_StoreAdaptiveChannelNeon(AecmCore* aecm,
const uint16_t* far_spectrum,
int32_t* echo_est) {
assert((uintptr_t)echo_est % 32 == 0);
assert((uintptr_t)(aecm->channelStored) % 16 == 0);
assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0);
// This is C code of following optimized code.
// During startup we store the channel every block.
// memcpy(aecm->channelStored,
// aecm->channelAdapt16,
// sizeof(int16_t) * PART_LEN1);
// Recalculate echo estimate
// for (i = 0; i < PART_LEN; i += 4) {
// echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i],
// far_spectrum[i]);
// echo_est[i + 1] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 1],
// far_spectrum[i + 1]);
// echo_est[i + 2] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 2],
// far_spectrum[i + 2]);
// echo_est[i + 3] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 3],
// far_spectrum[i + 3]);
// }
// echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i],
// far_spectrum[i]);
const uint16_t* far_spectrum_p = far_spectrum;
int16_t* start_adapt_p = aecm->channelAdapt16;
int16_t* start_stored_p = aecm->channelStored;
const int16_t* end_stored_p = aecm->channelStored + PART_LEN;
int32_t* echo_est_p = echo_est;
uint16x8_t far_spectrum_v;
int16x8_t adapt_v;
uint32x4_t echo_est_v_low, echo_est_v_high;
while (start_stored_p < end_stored_p) {
far_spectrum_v = vld1q_u16(far_spectrum_p);
adapt_v = vld1q_s16(start_adapt_p);
vst1q_s16(start_stored_p, adapt_v);
echo_est_v_low = vmull_u16(vget_low_u16(far_spectrum_v),
vget_low_u16(vreinterpretq_u16_s16(adapt_v)));
echo_est_v_high = vmull_u16(vget_high_u16(far_spectrum_v),
vget_high_u16(vreinterpretq_u16_s16(adapt_v)));
vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low));
vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high));
far_spectrum_p += 8;
start_adapt_p += 8;
start_stored_p += 8;
echo_est_p += 8;
}
aecm->channelStored[PART_LEN] = aecm->channelAdapt16[PART_LEN];
echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN],
far_spectrum[PART_LEN]);
}
static void SharpYUVFilterRow_NEON(const int16_t* A, const int16_t* B, int len,
const uint16_t* best_y, uint16_t* out) {
int i;
const int16x8_t max = vdupq_n_s16(MAX_Y);
const int16x8_t zero = vdupq_n_s16(0);
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t a0 = vld1q_s16(A + i + 0);
const int16x8_t a1 = vld1q_s16(A + i + 1);
const int16x8_t b0 = vld1q_s16(B + i + 0);
const int16x8_t b1 = vld1q_s16(B + i + 1);
const int16x8_t a0b1 = vaddq_s16(a0, b1);
const int16x8_t a1b0 = vaddq_s16(a1, b0);
const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0); // A0+A1+B0+B1
const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1); // 2*(A0+B1)
const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0); // 2*(A1+B0)
const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3);
const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3);
const int16x8_t d0 = vaddq_s16(c1, a0);
const int16x8_t d1 = vaddq_s16(c0, a1);
const int16x8_t e0 = vrshrq_n_s16(d0, 1);
const int16x8_t e1 = vrshrq_n_s16(d1, 1);
const int16x8x2_t f = vzipq_s16(e0, e1);
const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0));
const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8));
const int16x8_t h0 = vaddq_s16(g0, f.val[0]);
const int16x8_t h1 = vaddq_s16(g1, f.val[1]);
const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero);
const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero);
vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0));
vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1));
}
for (; i < len; ++i) {
const int a0b1 = A[i + 0] + B[i + 1];
const int a1b0 = A[i + 1] + B[i + 0];
const int a0a1b0b1 = a0b1 + a1b0 + 8;
const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0);
out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1);
}
}
static INLINE void highbd_idct32x32_1_add_neg_kernel(uint16_t **dest,
const int stride,
const int16x8_t res) {
const uint16x8_t a0 = vld1q_u16(*dest);
const uint16x8_t a1 = vld1q_u16(*dest + 8);
const uint16x8_t a2 = vld1q_u16(*dest + 16);
const uint16x8_t a3 = vld1q_u16(*dest + 24);
const int16x8_t b0 = vaddq_s16(res, vreinterpretq_s16_u16(a0));
const int16x8_t b1 = vaddq_s16(res, vreinterpretq_s16_u16(a1));
const int16x8_t b2 = vaddq_s16(res, vreinterpretq_s16_u16(a2));
const int16x8_t b3 = vaddq_s16(res, vreinterpretq_s16_u16(a3));
const uint16x8_t c0 = vqshluq_n_s16(b0, 0);
const uint16x8_t c1 = vqshluq_n_s16(b1, 0);
const uint16x8_t c2 = vqshluq_n_s16(b2, 0);
const uint16x8_t c3 = vqshluq_n_s16(b3, 0);
vst1q_u16(*dest, c0);
vst1q_u16(*dest + 8, c1);
vst1q_u16(*dest + 16, c2);
vst1q_u16(*dest + 24, c3);
*dest += stride;
}
static INLINE void highbd_idct32x32_1_add_pos_kernel(uint16_t **dest,
const int stride,
const int16x8_t res,
const int16x8_t max) {
const uint16x8_t a0 = vld1q_u16(*dest);
const uint16x8_t a1 = vld1q_u16(*dest + 8);
const uint16x8_t a2 = vld1q_u16(*dest + 16);
const uint16x8_t a3 = vld1q_u16(*dest + 24);
const int16x8_t b0 = vaddq_s16(res, vreinterpretq_s16_u16(a0));
const int16x8_t b1 = vaddq_s16(res, vreinterpretq_s16_u16(a1));
const int16x8_t b2 = vaddq_s16(res, vreinterpretq_s16_u16(a2));
const int16x8_t b3 = vaddq_s16(res, vreinterpretq_s16_u16(a3));
const int16x8_t c0 = vminq_s16(b0, max);
const int16x8_t c1 = vminq_s16(b1, max);
const int16x8_t c2 = vminq_s16(b2, max);
const int16x8_t c3 = vminq_s16(b3, max);
vst1q_u16(*dest, vreinterpretq_u16_s16(c0));
vst1q_u16(*dest + 8, vreinterpretq_u16_s16(c1));
vst1q_u16(*dest + 16, vreinterpretq_u16_s16(c2));
vst1q_u16(*dest + 24, vreinterpretq_u16_s16(c3));
*dest += stride;
}
/* u16x8 saturated sub */
void mw_neon_mm_qsub_u16x8(unsigned short * A, int Row, int Col, unsigned short * B, unsigned short * C)
{
uint16x8_t neon_a, neon_b, neon_c;
int size = Row * Col;
int i = 0;
int k = 0;
for (i = 8; i <= size ; i+=8)
{
k = i - 8;
neon_a = vld1q_u16(A + k);
neon_b = vld1q_u16(B + k);
neon_c = vqsubq_u16(neon_a, neon_b);
vst1q_u16(C + k, neon_c);
}
k = i - 8;
for (i = 0; i < size % 8; i++)
{
C[k + i] = A[k + i] + B[k + i];
}
}
void SkBlitLCD16OpaqueRow_neon(SkPMColor dst[], const uint16_t src[],
SkColor color, int width,
SkPMColor opaqueDst) {
int colR = SkColorGetR(color);
int colG = SkColorGetG(color);
int colB = SkColorGetB(color);
uint8x8_t vcolR, vcolG, vcolB;
uint8x8_t vopqDstA, vopqDstR, vopqDstG, vopqDstB;
if (width >= 8) {
vcolR = vdup_n_u8(colR);
vcolG = vdup_n_u8(colG);
vcolB = vdup_n_u8(colB);
vopqDstA = vdup_n_u8(SkGetPackedA32(opaqueDst));
vopqDstR = vdup_n_u8(SkGetPackedR32(opaqueDst));
vopqDstG = vdup_n_u8(SkGetPackedG32(opaqueDst));
vopqDstB = vdup_n_u8(SkGetPackedB32(opaqueDst));
}
while (width >= 8) {
uint8x8x4_t vdst;
uint16x8_t vmask;
uint16x8_t vmaskR, vmaskG, vmaskB;
uint8x8_t vsel_trans, vsel_opq;
vdst = vld4_u8((uint8_t*)dst);
vmask = vld1q_u16(src);
// Prepare compare masks
vsel_trans = vmovn_u16(vceqq_u16(vmask, vdupq_n_u16(0)));
vsel_opq = vmovn_u16(vceqq_u16(vmask, vdupq_n_u16(0xFFFF)));
// Get all the color masks on 5 bits
vmaskR = vshrq_n_u16(vmask, SK_R16_SHIFT);
vmaskG = vshrq_n_u16(vshlq_n_u16(vmask, SK_R16_BITS),
SK_B16_BITS + SK_R16_BITS + 1);
vmaskB = vmask & vdupq_n_u16(SK_B16_MASK);
// Upscale to 0..32
vmaskR = vmaskR + vshrq_n_u16(vmaskR, 4);
vmaskG = vmaskG + vshrq_n_u16(vmaskG, 4);
vmaskB = vmaskB + vshrq_n_u16(vmaskB, 4);
vdst.val[NEON_A] = vbsl_u8(vsel_trans, vdst.val[NEON_A], vdup_n_u8(0xFF));
vdst.val[NEON_A] = vbsl_u8(vsel_opq, vopqDstA, vdst.val[NEON_A]);
vdst.val[NEON_R] = SkBlend32_neon8(vcolR, vdst.val[NEON_R], vmaskR);
vdst.val[NEON_G] = SkBlend32_neon8(vcolG, vdst.val[NEON_G], vmaskG);
vdst.val[NEON_B] = SkBlend32_neon8(vcolB, vdst.val[NEON_B], vmaskB);
vdst.val[NEON_R] = vbsl_u8(vsel_opq, vopqDstR, vdst.val[NEON_R]);
vdst.val[NEON_G] = vbsl_u8(vsel_opq, vopqDstG, vdst.val[NEON_G]);
vdst.val[NEON_B] = vbsl_u8(vsel_opq, vopqDstB, vdst.val[NEON_B]);
vst4_u8((uint8_t*)dst, vdst);
dst += 8;
src += 8;
width -= 8;
}
// Leftovers
for (int i = 0; i < width; i++) {
dst[i] = SkBlendLCD16Opaque(colR, colG, colB, dst[i], src[i],
opaqueDst);
}
}
XnStatus XnPacked11DepthProcessor::Unpack11to16(const XnUInt8* pcInput, const XnUInt32 nInputSize, XnUInt32* pnActualRead)
{
const XnUInt8* pOrigInput = pcInput;
XnUInt32 nElements = nInputSize / XN_INPUT_ELEMENT_SIZE; // floored
XnUInt32 nNeededOutput = nElements * XN_OUTPUT_ELEMENT_SIZE;
*pnActualRead = 0;
XnBuffer* pWriteBuffer = GetWriteBuffer();
// Check there is enough room for the depth pixels
if (!CheckDepthBufferForOverflow(nNeededOutput))
{
return XN_STATUS_OUTPUT_BUFFER_OVERFLOW;
}
XnUInt16* pShiftOut = GetShiftsOutputBuffer();
XnUInt16* pnOutput = GetDepthOutputBuffer();
XnUInt16 a0,a1,a2,a3,a4,a5,a6,a7;
#ifdef XN_NEON
XnUInt16 shift[8];
XnUInt16 depth[8];
uint16x8_t Q0;
#endif
// Convert the 11bit packed data into 16bit shorts
for (XnUInt32 nElem = 0; nElem < nElements; ++nElem)
{
// input: 0, 1, 2,3, 4, 5, 6,7, 8, 9,10
// -,---,---,-,---,---,---,-,---,---,-
// bits: 8,3,5,6,2,8,1,7,4,4,7,1,8,2,6,5,3,8
// ---,---,-----,---,---,-----,---,---
// output: 0, 1, 2, 3, 4, 5, 6, 7
a0 = (XN_TAKE_BITS(pcInput[0],8,0) << 3) | XN_TAKE_BITS(pcInput[1],3,5);
a1 = (XN_TAKE_BITS(pcInput[1],5,0) << 6) | XN_TAKE_BITS(pcInput[2],6,2);
a2 = (XN_TAKE_BITS(pcInput[2],2,0) << 9) | (XN_TAKE_BITS(pcInput[3],8,0) << 1) | XN_TAKE_BITS(pcInput[4],1,7);
a3 = (XN_TAKE_BITS(pcInput[4],7,0) << 4) | XN_TAKE_BITS(pcInput[5],4,4);
a4 = (XN_TAKE_BITS(pcInput[5],4,0) << 7) | XN_TAKE_BITS(pcInput[6],7,1);
a5 = (XN_TAKE_BITS(pcInput[6],1,0) << 10) | (XN_TAKE_BITS(pcInput[7],8,0) << 2) | XN_TAKE_BITS(pcInput[8],2,6);
a6 = (XN_TAKE_BITS(pcInput[8],6,0) << 5) | XN_TAKE_BITS(pcInput[9],5,3);
a7 = (XN_TAKE_BITS(pcInput[9],3,0) << 8) | XN_TAKE_BITS(pcInput[10],8,0);
#ifdef XN_NEON
shift[0] = (((a0) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a0) : 0);
shift[1] = (((a1) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a1) : 0);
shift[2] = (((a2) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a2) : 0);
shift[3] = (((a3) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a3) : 0);
shift[4] = (((a4) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a4) : 0);
shift[5] = (((a5) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a5) : 0);
shift[6] = (((a6) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a6) : 0);
shift[7] = (((a7) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a7) : 0);
depth[0] = GetOutput(a0);
depth[1] = GetOutput(a1);
depth[2] = GetOutput(a2);
depth[3] = GetOutput(a3);
depth[4] = GetOutput(a4);
depth[5] = GetOutput(a5);
depth[6] = GetOutput(a6);
depth[7] = GetOutput(a7);
// Load
Q0 = vld1q_u16(depth);
// Store
vst1q_u16(pnOutput, Q0);
// Load
Q0 = vld1q_u16(shift);
// Store
vst1q_u16(pShiftOut, Q0);
#else
pShiftOut[0] = (((a0) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a0) : 0);
pShiftOut[1] = (((a1) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a1) : 0);
pShiftOut[2] = (((a2) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a2) : 0);
pShiftOut[3] = (((a3) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a3) : 0);
pShiftOut[4] = (((a4) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a4) : 0);
pShiftOut[5] = (((a5) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a5) : 0);
pShiftOut[6] = (((a6) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a6) : 0);
pShiftOut[7] = (((a7) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (a7) : 0);
pnOutput[0] = GetOutput(a0);
pnOutput[1] = GetOutput(a1);
pnOutput[2] = GetOutput(a2);
pnOutput[3] = GetOutput(a3);
pnOutput[4] = GetOutput(a4);
pnOutput[5] = GetOutput(a5);
pnOutput[6] = GetOutput(a6);
pnOutput[7] = GetOutput(a7);
#endif
pcInput += XN_INPUT_ELEMENT_SIZE;
pnOutput += 8;
pShiftOut += 8;
}
*pnActualRead = (XnUInt32)(pcInput - pOrigInput);
pWriteBuffer->UnsafeUpdateSize(nNeededOutput);
return XN_STATUS_OK;
}
XnStatus XnPacked12DepthProcessor::Unpack12to16(const XnUInt8* pcInput, const XnUInt32 nInputSize, XnUInt32* pnActualRead)
{
const XnUInt8* pOrigInput = pcInput;
XnUInt32 nElements = nInputSize / XN_INPUT_ELEMENT_SIZE; // floored
XnUInt32 nNeededOutput = nElements * XN_OUTPUT_ELEMENT_SIZE;
*pnActualRead = 0;
XnBuffer* pWriteBuffer = GetWriteBuffer();
if (!CheckDepthBufferForOverflow(nNeededOutput))
{
return XN_STATUS_OUTPUT_BUFFER_OVERFLOW;
}
XnUInt16* pnOutput = GetDepthOutputBuffer();
XnUInt16* pShiftOut = GetShiftsOutputBuffer();
XnUInt16 shift[16];
#ifdef XN_NEON
XnUInt16 depth[16];
uint8x8x3_t inD3;
uint8x8_t rshft4D, lshft4D;
uint16x8_t rshft4Q, lshft4Q;
uint16x8_t depthQ;
uint16x8x2_t shiftQ2;
#endif
// Convert the 11bit packed data into 16bit shorts
for (XnUInt32 nElem = 0; nElem < nElements; ++nElem)
{
#ifndef XN_NEON
// input: 0, 1,2,3, 4,5,6, 7,8,9, 10,11,12, 13,14,15, 16,17,18, 19,20,21, 22,23
// -,---,-,-,---,-,-,---,-,-,---,--,--,---,--,--,---,--,--,---,--,--,---,--
// bits: 8,4,4,8,8,4,4,8,8,4,4,8,8,4,4, 8, 8,4,4, 8, 8,4,4, 8, 8,4,4, 8, 8,4,4, 8
// ---,---,---,---,---,---,---,----,----,----,----,----,----,----,----,----
// output: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
shift[0] = (XN_TAKE_BITS(pcInput[0],8,0) << 4) | XN_TAKE_BITS(pcInput[1],4,4);
shift[1] = (XN_TAKE_BITS(pcInput[1],4,0) << 8) | XN_TAKE_BITS(pcInput[2],8,0);
shift[2] = (XN_TAKE_BITS(pcInput[3],8,0) << 4) | XN_TAKE_BITS(pcInput[4],4,4);
shift[3] = (XN_TAKE_BITS(pcInput[4],4,0) << 8) | XN_TAKE_BITS(pcInput[5],8,0);
shift[4] = (XN_TAKE_BITS(pcInput[6],8,0) << 4) | XN_TAKE_BITS(pcInput[7],4,4);
shift[5] = (XN_TAKE_BITS(pcInput[7],4,0) << 8) | XN_TAKE_BITS(pcInput[8],8,0);
shift[6] = (XN_TAKE_BITS(pcInput[9],8,0) << 4) | XN_TAKE_BITS(pcInput[10],4,4);
shift[7] = (XN_TAKE_BITS(pcInput[10],4,0) << 8) | XN_TAKE_BITS(pcInput[11],8,0);
shift[8] = (XN_TAKE_BITS(pcInput[12],8,0) << 4) | XN_TAKE_BITS(pcInput[13],4,4);
shift[9] = (XN_TAKE_BITS(pcInput[13],4,0) << 8) | XN_TAKE_BITS(pcInput[14],8,0);
shift[10] = (XN_TAKE_BITS(pcInput[15],8,0) << 4) | XN_TAKE_BITS(pcInput[16],4,4);
shift[11] = (XN_TAKE_BITS(pcInput[16],4,0) << 8) | XN_TAKE_BITS(pcInput[17],8,0);
shift[12] = (XN_TAKE_BITS(pcInput[18],8,0) << 4) | XN_TAKE_BITS(pcInput[19],4,4);
shift[13] = (XN_TAKE_BITS(pcInput[19],4,0) << 8) | XN_TAKE_BITS(pcInput[20],8,0);
shift[14] = (XN_TAKE_BITS(pcInput[21],8,0) << 4) | XN_TAKE_BITS(pcInput[22],4,4);
shift[15] = (XN_TAKE_BITS(pcInput[22],4,0) << 8) | XN_TAKE_BITS(pcInput[23],8,0);
pShiftOut[0] = (((shift[0]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[0]) : 0);
pShiftOut[1] = (((shift[1]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[1]) : 0);
pShiftOut[2] = (((shift[2]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[2]) : 0);
pShiftOut[3] = (((shift[3]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[3]) : 0);
pShiftOut[4] = (((shift[4]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[4]) : 0);
pShiftOut[5] = (((shift[5]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[5]) : 0);
pShiftOut[6] = (((shift[6]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[6]) : 0);
pShiftOut[7] = (((shift[7]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[7]) : 0);
pShiftOut[8] = (((shift[0]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[8]) : 0);
pShiftOut[9] = (((shift[1]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[9]) : 0);
pShiftOut[10] = (((shift[2]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[10]) : 0);
pShiftOut[11] = (((shift[3]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[11]) : 0);
pShiftOut[12] = (((shift[4]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[12]) : 0);
pShiftOut[13] = (((shift[5]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[13]) : 0);
pShiftOut[14] = (((shift[6]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[14]) : 0);
pShiftOut[15] = (((shift[7]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[15]) : 0);
pnOutput[0] = GetOutput(shift[0]);
pnOutput[1] = GetOutput(shift[1]);
pnOutput[2] = GetOutput(shift[2]);
pnOutput[3] = GetOutput(shift[3]);
pnOutput[4] = GetOutput(shift[4]);
pnOutput[5] = GetOutput(shift[5]);
pnOutput[6] = GetOutput(shift[6]);
pnOutput[7] = GetOutput(shift[7]);
pnOutput[8] = GetOutput(shift[8]);
pnOutput[9] = GetOutput(shift[9]);
pnOutput[10] = GetOutput(shift[10]);
pnOutput[11] = GetOutput(shift[11]);
pnOutput[12] = GetOutput(shift[12]);
pnOutput[13] = GetOutput(shift[13]);
pnOutput[14] = GetOutput(shift[14]);
pnOutput[15] = GetOutput(shift[15]);
#else
// input: 0, 1,2 (X8)
// -,---,-
// bits: 8,4,4,8 (X8)
// ---,---
// output: 0, 1 (X8)
// Split 24 bytes into 3 vectors (64 bit each)
inD3 = vld3_u8(pcInput);
// rshft4D0 contains 4 MSB of second vector (placed at offset 0)
rshft4D = vshr_n_u8(inD3.val[1], 4);
// lshft4D0 contains 4 LSB of second vector (placed at offset 4)
lshft4D = vshl_n_u8(inD3.val[1], 4);
// Expand 64 bit vectors to 128 bit (8 values of 16 bits)
shiftQ2.val[0] = vmovl_u8(inD3.val[0]);
shiftQ2.val[1] = vmovl_u8(inD3.val[2]);
rshft4Q = vmovl_u8(rshft4D);
lshft4Q = vmovl_u8(lshft4D);
// Even indexed shift = 8 bits from first vector + 4 MSB bits of second vector
shiftQ2.val[0] = vshlq_n_u16(shiftQ2.val[0], 4);
shiftQ2.val[0] = vorrq_u16(shiftQ2.val[0], rshft4Q);
// Odd indexed shift = 4 LSB bits of second vector + 8 bits from third vector
lshft4Q = vshlq_n_u16(lshft4Q, 4);
shiftQ2.val[1] = vorrq_u16(shiftQ2.val[1], lshft4Q);
// Interleave shift values to a single vector
vst2q_u16(shift, shiftQ2);
shift[0] = (((shift[0]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[0]) : 0);
shift[1] = (((shift[1]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[1]) : 0);
shift[2] = (((shift[2]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[2]) : 0);
shift[3] = (((shift[3]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[3]) : 0);
shift[4] = (((shift[4]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[4]) : 0);
shift[5] = (((shift[5]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[5]) : 0);
shift[6] = (((shift[6]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[6]) : 0);
shift[7] = (((shift[7]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[7]) : 0);
shift[8] = (((shift[0]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[8]) : 0);
shift[9] = (((shift[1]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[9]) : 0);
shift[10] = (((shift[2]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[10]) : 0);
shift[11] = (((shift[3]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[11]) : 0);
shift[12] = (((shift[4]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[12]) : 0);
shift[13] = (((shift[5]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[13]) : 0);
shift[14] = (((shift[6]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[14]) : 0);
shift[15] = (((shift[7]) < (XN_DEVICE_SENSOR_MAX_SHIFT_VALUE-1)) ? (shift[15]) : 0);
depth[0] = GetOutput(shift[0]);
depth[1] = GetOutput(shift[1]);
depth[2] = GetOutput(shift[2]);
depth[3] = GetOutput(shift[3]);
depth[4] = GetOutput(shift[4]);
depth[5] = GetOutput(shift[5]);
depth[6] = GetOutput(shift[6]);
depth[7] = GetOutput(shift[7]);
// Load
depthQ = vld1q_u16(depth);
//Store
vst1q_u16(pnOutput, depthQ);
// Load
depthQ = vld1q_u16(shift);
// Store
vst1q_u16(pShiftOut, depthQ);
depth[8] = GetOutput(shift[8]);
depth[9] = GetOutput(shift[9]);
depth[10] = GetOutput(shift[10]);
depth[11] = GetOutput(shift[11]);
depth[12] = GetOutput(shift[12]);
depth[13] = GetOutput(shift[13]);
depth[14] = GetOutput(shift[14]);
depth[15] = GetOutput(shift[15]);
// Load
depthQ = vld1q_u16(depth + 8);
// Store
vst1q_u16(pnOutput + 8, depthQ);
// Load
depthQ = vld1q_u16(shift + 8);
// Store
vst1q_u16(pShiftOut + 8, depthQ);
#endif
pcInput += XN_INPUT_ELEMENT_SIZE;
pnOutput += 16;
pShiftOut += 16;
}
*pnActualRead = (XnUInt32)(pcInput - pOrigInput);
pWriteBuffer->UnsafeUpdateSize(nNeededOutput);
return XN_STATUS_OK;
}
void test_vld1Qu16 (void)
{
uint16x8_t out_uint16x8_t;
out_uint16x8_t = vld1q_u16 (0);
}
void meanStdDev(const Size2D &size,
const u16 * srcBase, ptrdiff_t srcStride,
f32 * pMean, f32 * pStdDev)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
size_t blockSize0 = 1 << 10, roiw4 = size.width & ~3;
f64 fsum = 0.0f, fsqsum = 0.0f;
f32 arsum[8];
uint32x4_t v_zero = vdupq_n_u32(0u), v_sum;
float32x4_t v_zero_f = vdupq_n_f32(0.0f), v_sqsum;
for (size_t i = 0; i < size.height; ++i)
{
const u16 * src = internal::getRowPtr(srcBase, srcStride, i);
size_t j = 0u;
while (j < roiw4)
{
size_t blockSize = std::min(roiw4 - j, blockSize0) + j;
v_sum = v_zero;
v_sqsum = v_zero_f;
for ( ; j + 16 < blockSize ; j += 16)
{
internal::prefetch(src + j);
uint16x8_t v_src0 = vld1q_u16(src + j), v_src1 = vld1q_u16(src + j + 8);
// 0
uint32x4_t v_srclo = vmovl_u16(vget_low_u16(v_src0));
uint32x4_t v_srchi = vmovl_u16(vget_high_u16(v_src0));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
float32x4_t v_srclo_f = vcvtq_f32_u32(v_srclo);
float32x4_t v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
// 1
v_srclo = vmovl_u16(vget_low_u16(v_src1));
v_srchi = vmovl_u16(vget_high_u16(v_src1));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
v_srclo_f = vcvtq_f32_u32(v_srclo);
v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
}
for ( ; j < blockSize; j += 4)
{
uint32x4_t v_src = vmovl_u16(vld1_u16(src + j));
float32x4_t v_src_f = vcvtq_f32_u32(v_src);
v_sum = vaddq_u32(v_sum, v_src);
v_sqsum = vmlaq_f32(v_sqsum, v_src_f, v_src_f);
}
vst1q_f32(arsum, vcvtq_f32_u32(v_sum));
vst1q_f32(arsum + 4, v_sqsum);
fsum += (f64)arsum[0] + arsum[1] + arsum[2] + arsum[3];
fsqsum += (f64)arsum[4] + arsum[5] + arsum[6] + arsum[7];
}
// collect a few last elements in the current row
for ( ; j < size.width; ++j)
{
f32 srcval = src[j];
fsum += srcval;
fsqsum += srcval * srcval;
}
}
// calc mean and stddev
f64 itotal = 1.0 / size.total();
f64 mean = fsum * itotal;
f64 stddev = sqrt(std::max(fsqsum * itotal - mean * mean, 0.0));
if (pMean)
*pMean = mean;
if (pStdDev)
*pStdDev = stddev;
#else
(void)size;
(void)srcBase;
(void)srcStride;
(void)pMean;
(void)pStdDev;
#endif
}
void vpx_highbd_convolve_avg_neon(const uint8_t *src8, ptrdiff_t src_stride,
uint8_t *dst8, ptrdiff_t dst_stride,
const int16_t *filter_x, int filter_x_stride,
const int16_t *filter_y, int filter_y_stride,
int w, int h, int bd) {
const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
(void)filter_x;
(void)filter_x_stride;
(void)filter_y;
(void)filter_y_stride;
(void)bd;
if (w < 8) { // avg4
uint16x4_t s0, s1, d0, d1;
uint16x8_t s01, d01;
do {
s0 = vld1_u16(src);
d0 = vld1_u16(dst);
src += src_stride;
s1 = vld1_u16(src);
d1 = vld1_u16(dst + dst_stride);
src += src_stride;
s01 = vcombine_u16(s0, s1);
d01 = vcombine_u16(d0, d1);
d01 = vrhaddq_u16(s01, d01);
vst1_u16(dst, vget_low_u16(d01));
dst += dst_stride;
vst1_u16(dst, vget_high_u16(d01));
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w == 8) { // avg8
uint16x8_t s0, s1, d0, d1;
do {
s0 = vld1q_u16(src);
d0 = vld1q_u16(dst);
src += src_stride;
s1 = vld1q_u16(src);
d1 = vld1q_u16(dst + dst_stride);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
vst1q_u16(dst, d0);
dst += dst_stride;
vst1q_u16(dst, d1);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w < 32) { // avg16
uint16x8_t s0l, s0h, s1l, s1h, d0l, d0h, d1l, d1h;
do {
s0l = vld1q_u16(src);
s0h = vld1q_u16(src + 8);
d0l = vld1q_u16(dst);
d0h = vld1q_u16(dst + 8);
src += src_stride;
s1l = vld1q_u16(src);
s1h = vld1q_u16(src + 8);
d1l = vld1q_u16(dst + dst_stride);
d1h = vld1q_u16(dst + dst_stride + 8);
src += src_stride;
d0l = vrhaddq_u16(s0l, d0l);
d0h = vrhaddq_u16(s0h, d0h);
d1l = vrhaddq_u16(s1l, d1l);
d1h = vrhaddq_u16(s1h, d1h);
vst1q_u16(dst, d0l);
vst1q_u16(dst + 8, d0h);
dst += dst_stride;
vst1q_u16(dst, d1l);
vst1q_u16(dst + 8, d1h);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w == 32) { // avg32
uint16x8_t s0, s1, s2, s3, d0, d1, d2, d3;
do {
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
dst += dst_stride;
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else { // avg64
uint16x8_t s0, s1, s2, s3, d0, d1, d2, d3;
do {
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
s0 = vld1q_u16(src + 32);
s1 = vld1q_u16(src + 40);
s2 = vld1q_u16(src + 48);
s3 = vld1q_u16(src + 56);
d0 = vld1q_u16(dst + 32);
d1 = vld1q_u16(dst + 40);
d2 = vld1q_u16(dst + 48);
d3 = vld1q_u16(dst + 56);
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst + 32, d0);
vst1q_u16(dst + 40, d1);
vst1q_u16(dst + 48, d2);
vst1q_u16(dst + 56, d3);
src += src_stride;
dst += dst_stride;
} while (--h);
}
}
XnStatus Link12BitS2DParser::Unpack12to16(const XnUInt8* pcInput,XnUInt8* pDest, const XnUInt32 nInputSize, XnUInt32* pnActualRead, XnUInt32* pnActualWritten)
{
const XnUInt8* pOrigInput = (XnUInt8*)pcInput;
XnUInt32 nElements = nInputSize / XN_INPUT_ELEMENT_SIZE; // floored
//XnUInt32 nNeededOutput = nElements * XN_OUTPUT_ELEMENT_SIZE;
*pnActualRead = 0;
XnUInt16 *pnOutput = (XnUInt16*)pDest;
XnUInt16 shift[16];
#ifdef XN_NEON
XnUInt16 depth[16];
uint8x8x3_t inD3;
uint8x8_t rshft4D, lshft4D;
uint16x8_t rshft4Q, lshft4Q;
uint16x8_t depthQ;
uint16x8x2_t shiftQ2;
#endif
// Convert the 11bit packed data into 16bit shorts
for (XnUInt32 nElem = 0; nElem < nElements; ++nElem)
{
#ifndef XN_NEON
// input: 0, 1,2,3, 4,5,6, 7,8,9, 10,11,12, 13,14,15, 16,17,18, 19,20,21, 22,23
// -,---,-,-,---,-,-,---,-,-,---,--,--,---,--,--,---,--,--,---,--,--,---,--
// bits: 8,4,4,8,8,4,4,8,8,4,4,8,8,4,4, 8, 8,4,4, 8, 8,4,4, 8, 8,4,4, 8, 8,4,4, 8
// ---,---,---,---,---,---,---,----,----,----,----,----,----,----,----,----
// output: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
shift[0] = (XN_TAKE_BITS(pcInput[0],8,0) << 4) | XN_TAKE_BITS(pcInput[1],4,4);
shift[1] = (XN_TAKE_BITS(pcInput[1],4,0) << 8) | XN_TAKE_BITS(pcInput[2],8,0);
shift[2] = (XN_TAKE_BITS(pcInput[3],8,0) << 4) | XN_TAKE_BITS(pcInput[4],4,4);
shift[3] = (XN_TAKE_BITS(pcInput[4],4,0) << 8) | XN_TAKE_BITS(pcInput[5],8,0);
shift[4] = (XN_TAKE_BITS(pcInput[6],8,0) << 4) | XN_TAKE_BITS(pcInput[7],4,4);
shift[5] = (XN_TAKE_BITS(pcInput[7],4,0) << 8) | XN_TAKE_BITS(pcInput[8],8,0);
shift[6] = (XN_TAKE_BITS(pcInput[9],8,0) << 4) | XN_TAKE_BITS(pcInput[10],4,4);
shift[7] = (XN_TAKE_BITS(pcInput[10],4,0) << 8) | XN_TAKE_BITS(pcInput[11],8,0);
shift[8] = (XN_TAKE_BITS(pcInput[12],8,0) << 4) | XN_TAKE_BITS(pcInput[13],4,4);
shift[9] = (XN_TAKE_BITS(pcInput[13],4,0) << 8) | XN_TAKE_BITS(pcInput[14],8,0);
shift[10] = (XN_TAKE_BITS(pcInput[15],8,0) << 4) | XN_TAKE_BITS(pcInput[16],4,4);
shift[11] = (XN_TAKE_BITS(pcInput[16],4,0) << 8) | XN_TAKE_BITS(pcInput[17],8,0);
shift[12] = (XN_TAKE_BITS(pcInput[18],8,0) << 4) | XN_TAKE_BITS(pcInput[19],4,4);
shift[13] = (XN_TAKE_BITS(pcInput[19],4,0) << 8) | XN_TAKE_BITS(pcInput[20],8,0);
shift[14] = (XN_TAKE_BITS(pcInput[21],8,0) << 4) | XN_TAKE_BITS(pcInput[22],4,4);
shift[15] = (XN_TAKE_BITS(pcInput[22],4,0) << 8) | XN_TAKE_BITS(pcInput[23],8,0);
pnOutput[0] = m_pShiftToDepth[(shift[0])];
pnOutput[1] = m_pShiftToDepth[(shift[1])];
pnOutput[2] = m_pShiftToDepth[(shift[2])];
pnOutput[3] = m_pShiftToDepth[(shift[3])];
pnOutput[4] = m_pShiftToDepth[(shift[4])];
pnOutput[5] = m_pShiftToDepth[(shift[5])];
pnOutput[6] = m_pShiftToDepth[(shift[6])];
pnOutput[7] = m_pShiftToDepth[(shift[7])];
pnOutput[8] = m_pShiftToDepth[(shift[8])];
pnOutput[9] = m_pShiftToDepth[(shift[9])];
pnOutput[10] = m_pShiftToDepth[(shift[10])];
pnOutput[11] = m_pShiftToDepth[(shift[11])];
pnOutput[12] = m_pShiftToDepth[(shift[12])];
pnOutput[13] = m_pShiftToDepth[(shift[13])];
pnOutput[14] = m_pShiftToDepth[(shift[14])];
pnOutput[15] = m_pShiftToDepth[(shift[15])];
#else
// input: 0, 1,2 (X8)
// -,---,-
// bits: 8,4,4,8 (X8)
// ---,---
// output: 0, 1 (X8)
// Split 24 bytes into 3 vectors (64 bit each)
inD3 = vld3_u8(pcInput);
// rshft4D0 contains 4 MSB of second vector (placed at offset 0)
rshft4D = vshr_n_u8(inD3.val[1], 4);
// lshft4D0 contains 4 LSB of second vector (placed at offset 4)
lshft4D = vshl_n_u8(inD3.val[1], 4);
// Expand 64 bit vectors to 128 bit (8 values of 16 bits)
shiftQ2.val[0] = vmovl_u8(inD3.val[0]);
shiftQ2.val[1] = vmovl_u8(inD3.val[2]);
rshft4Q = vmovl_u8(rshft4D);
lshft4Q = vmovl_u8(lshft4D);
// Even indexed shift = 8 bits from first vector + 4 MSB bits of second vector
shiftQ2.val[0] = vshlq_n_u16(shiftQ2.val[0], 4);
shiftQ2.val[0] = vorrq_u16(shiftQ2.val[0], rshft4Q);
// Odd indexed shift = 4 LSB bits of second vector + 8 bits from third vector
lshft4Q = vshlq_n_u16(lshft4Q, 4);
shiftQ2.val[1] = vorrq_u16(shiftQ2.val[1], lshft4Q);
// Interleave shift values to a single vector
vst2q_u16(shift, shiftQ2);
depth[0] = m_pShiftToDepth[(shift[0])];
depth[1] = m_pShiftToDepth[(shift[1])];
depth[2] = m_pShiftToDepth[(shift[2])];
depth[3] = m_pShiftToDepth[(shift[3])];
depth[4] = m_pShiftToDepth[(shift[4])];
depth[5] = m_pShiftToDepth[(shift[5])];
depth[6] = m_pShiftToDepth[(shift[6])];
depth[7] = m_pShiftToDepth[(shift[7])];
// Load
depthQ = vld1q_u16(depth);
//Store
vst1q_u16(pnOutput, depthQ);
depth[8] = m_pShiftToDepth[(shift[8])];
depth[9] = m_pShiftToDepth[(shift[9])];
depth[10] = m_pShiftToDepth[(shift[10])];
depth[11] = m_pShiftToDepth[(shift[11])];
depth[12] = m_pShiftToDepth[(shift[12])];
depth[13] = m_pShiftToDepth[(shift[13])];
depth[14] = m_pShiftToDepth[(shift[14])];
depth[15] = m_pShiftToDepth[(shift[15])];
// Load
depthQ = vld1q_u16(depth + 8);
// Store
vst1q_u16(pnOutput + 8, depthQ);
#endif
pcInput += XN_INPUT_ELEMENT_SIZE;
pnOutput += 16;
}
*pnActualRead = (XnUInt32)(pcInput - pOrigInput); // total bytes
*pnActualWritten = (XnUInt32)((XnUInt8*)pnOutput - pDest);
return XN_STATUS_OK;
}
XnStatus XnPacked11DepthProcessor::Unpack11to16(const XnUInt8* pcInput, const XnUInt32 nInputSize, XnUInt32* pnActualRead)
{
const XnUInt8* pOrigInput = pcInput;
XnUInt32 nElements = nInputSize / XN_INPUT_ELEMENT_SIZE; // floored
XnUInt32 nNeededOutput = nElements * XN_OUTPUT_ELEMENT_SIZE;
*pnActualRead = 0;
XnBuffer* pWriteBuffer = GetWriteBuffer();
// Check there is enough room for the depth pixels
if (!CheckWriteBufferForOverflow(nNeededOutput))
{
return XN_STATUS_OUTPUT_BUFFER_OVERFLOW;
}
XnUInt16* pnOutput = (XnUInt16*)pWriteBuffer->GetUnsafeWritePointer();
XnUInt16 a0,a1,a2,a3,a4,a5,a6,a7;
#ifdef XN_NEON
XnUInt16 depth[8];
uint16x8_t Q0;
#endif
// Convert the 11bit packed data into 16bit shorts
for (XnUInt32 nElem = 0; nElem < nElements; ++nElem)
{
if(m_nScaleFactor > 1)
{
XnUInt32 px = m_nOffsetInFrame%m_CurrentVideoMode.resolutionX;
XnUInt32 py = (m_nOffsetInFrame)/m_CurrentVideoMode.resolutionX;
if(py%m_nScaleFactor != 0)
{
// Skip as many pixels as possible
XnUInt32 nEltsToSkip =
XN_MIN(nElements - nElem,
(m_CurrentVideoMode.resolutionX - px)/8
+ (m_nScaleFactor-(py%m_nScaleFactor) - 1)*m_CurrentVideoMode.resolutionX/8);
// ::memset(pnOutput, 0, nEltsToSkip*8*sizeof(XnUInt16));
pcInput += nEltsToSkip*XN_INPUT_ELEMENT_SIZE;
pnOutput += nEltsToSkip*8;
m_nOffsetInFrame += nEltsToSkip*8;
nElem += (nEltsToSkip-1);
continue;
}
}
// input: 0, 1, 2,3, 4, 5, 6,7, 8, 9,10
// -,---,---,-,---,---,---,-,---,---,-
// bits: 8,3,5,6,2,8,1,7,4,4,7,1,8,2,6,5,3,8
// ---,---,-----,---,---,-----,---,---
// output: 0, 1, 2, 3, 4, 5, 6, 7
if(m_nScaleFactor == 2)
{
a0 = (XN_TAKE_BITS(pcInput[0],8,0) << 3) | XN_TAKE_BITS(pcInput[1],3,5);
a2 = (XN_TAKE_BITS(pcInput[2],2,0) << 9) | (XN_TAKE_BITS(pcInput[3],8,0) << 1) | XN_TAKE_BITS(pcInput[4],1,7);
a4 = (XN_TAKE_BITS(pcInput[5],4,0) << 7) | XN_TAKE_BITS(pcInput[6],7,1);
a6 = (XN_TAKE_BITS(pcInput[8],6,0) << 5) | XN_TAKE_BITS(pcInput[9],5,3);
}
else if(m_nScaleFactor == 4)
{
a0 = (XN_TAKE_BITS(pcInput[0],8,0) << 3) | XN_TAKE_BITS(pcInput[1],3,5);
a4 = (XN_TAKE_BITS(pcInput[5],4,0) << 7) | XN_TAKE_BITS(pcInput[6],7,1);
}
else
{
a0 = (XN_TAKE_BITS(pcInput[0],8,0) << 3) | XN_TAKE_BITS(pcInput[1],3,5);
a1 = (XN_TAKE_BITS(pcInput[1],5,0) << 6) | XN_TAKE_BITS(pcInput[2],6,2);
a2 = (XN_TAKE_BITS(pcInput[2],2,0) << 9) | (XN_TAKE_BITS(pcInput[3],8,0) << 1) | XN_TAKE_BITS(pcInput[4],1,7);
a3 = (XN_TAKE_BITS(pcInput[4],7,0) << 4) | XN_TAKE_BITS(pcInput[5],4,4);
a4 = (XN_TAKE_BITS(pcInput[5],4,0) << 7) | XN_TAKE_BITS(pcInput[6],7,1);
a5 = (XN_TAKE_BITS(pcInput[6],1,0) << 10) | (XN_TAKE_BITS(pcInput[7],8,0) << 2) | XN_TAKE_BITS(pcInput[8],2,6);
a6 = (XN_TAKE_BITS(pcInput[8],6,0) << 5) | XN_TAKE_BITS(pcInput[9],5,3);
a7 = (XN_TAKE_BITS(pcInput[9],3,0) << 8) | XN_TAKE_BITS(pcInput[10],8,0);
}
#ifdef XN_NEON
depth[0] = GetOutput(a0);
depth[1] = GetOutput(a1);
depth[2] = GetOutput(a2);
depth[3] = GetOutput(a3);
depth[4] = GetOutput(a4);
depth[5] = GetOutput(a5);
depth[6] = GetOutput(a6);
depth[7] = GetOutput(a7);
// Load
Q0 = vld1q_u16(depth);
// Store
vst1q_u16(pnOutput, Q0);
#else
if(m_nScaleFactor == 2)
{
*pnOutput++ = GetOutput(a0);
*pnOutput++ = 0;
*pnOutput++ = GetOutput(a2);
*pnOutput++ = 0;
*pnOutput++ = GetOutput(a4);
*pnOutput++ = 0;
*pnOutput++ = GetOutput(a6);
*pnOutput++ = 0;
}
else if(m_nScaleFactor == 4)
{
*pnOutput++ = GetOutput(a0);
*pnOutput++ = 0;
*pnOutput++ = 0;
*pnOutput++ = 0;
*pnOutput++ = GetOutput(a4);
*pnOutput++ = 0;
*pnOutput++ = 0;
*pnOutput++ = 0;
}
else
{
*pnOutput++ = GetOutput(a0);
*pnOutput++ = GetOutput(a1);
*pnOutput++ = GetOutput(a2);
*pnOutput++ = GetOutput(a3);
*pnOutput++ = GetOutput(a4);
*pnOutput++ = GetOutput(a5);
*pnOutput++ = GetOutput(a6);
*pnOutput++ = GetOutput(a7);
}
#endif
pcInput += XN_INPUT_ELEMENT_SIZE;
m_nOffsetInFrame+=8;
}
*pnActualRead = (XnUInt32)(pcInput - pOrigInput);
pWriteBuffer->UnsafeUpdateSize(nNeededOutput);
return XN_STATUS_OK;
}
/* u16x8 mm mul */
void mw_neon_mm_mul_u16x8(unsigned short * A, int Row, int T, unsigned short * B, int Col, unsigned short * C)
{
int i, k, j;
uint16x8_t neon_b, neon_c;
uint16x8_t neon_a0, neon_a1, neon_a2, neon_a3, neon_a4, neon_a5, neon_a6, neon_a7;
uint16x8_t neon_b0, neon_b1, neon_b2, neon_b3, neon_b4, neon_b5, neon_b6, neon_b7;
for (i = 0; i < Row; i+=8)
{
for (k = 0; k < Col; k+=1)
{
neon_c = vmovq_n_u16(0);
for (j = 0; j < T; j+=8)
{
int j_T = j * T + i;
int k_Row = k * Row;
neon_a0 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a1 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a2 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a3 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a4 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a5 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a6 = vld1q_u16(A + j_T);
j_T+=Row;
neon_a7 = vld1q_u16(A + j_T);
neon_b = vld1q_u16(B + k_Row + j);
neon_b0 = vdupq_n_u16(vgetq_lane_u16(neon_b, 0));
neon_b1 = vdupq_n_u16(vgetq_lane_u16(neon_b, 1));
neon_b2 = vdupq_n_u16(vgetq_lane_u16(neon_b, 2));
neon_b3 = vdupq_n_u16(vgetq_lane_u16(neon_b, 3));
neon_b4 = vdupq_n_u16(vgetq_lane_u16(neon_b, 4));
neon_b5 = vdupq_n_u16(vgetq_lane_u16(neon_b, 5));
neon_b6 = vdupq_n_u16(vgetq_lane_u16(neon_b, 6));
neon_b7 = vdupq_n_u16(vgetq_lane_u16(neon_b, 7));
neon_c = vaddq_u16(vmulq_u16(neon_a0, neon_b0), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a1, neon_b1), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a2, neon_b2), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a3, neon_b3), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a4, neon_b4), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a5, neon_b5), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a6, neon_b6), neon_c);
neon_c = vaddq_u16(vmulq_u16(neon_a7, neon_b7), neon_c);
vst1q_lane_u16(C + k_Row + i, neon_c, 0);
vst1q_lane_u16(C + k_Row + i + 1, neon_c, 1);
vst1q_lane_u16(C + k_Row + i + 2, neon_c, 2);
vst1q_lane_u16(C + k_Row + i + 3, neon_c, 3);
vst1q_lane_u16(C + k_Row + i + 4, neon_c, 4);
vst1q_lane_u16(C + k_Row + i + 5, neon_c, 5);
vst1q_lane_u16(C + k_Row + i + 6, neon_c, 6);
vst1q_lane_u16(C + k_Row + i + 7, neon_c, 7);
}
}
}
}
void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore* aecm,
const uint16_t* far_spectrum,
int32_t* echo_est,
uint32_t* far_energy,
uint32_t* echo_energy_adapt,
uint32_t* echo_energy_stored) {
int16_t* start_stored_p = aecm->channelStored;
int16_t* start_adapt_p = aecm->channelAdapt16;
int32_t* echo_est_p = echo_est;
const int16_t* end_stored_p = aecm->channelStored + PART_LEN;
const uint16_t* far_spectrum_p = far_spectrum;
int16x8_t store_v, adapt_v;
uint16x8_t spectrum_v;
uint32x4_t echo_est_v_low, echo_est_v_high;
uint32x4_t far_energy_v, echo_stored_v, echo_adapt_v;
far_energy_v = vdupq_n_u32(0);
echo_adapt_v = vdupq_n_u32(0);
echo_stored_v = vdupq_n_u32(0);
// Get energy for the delayed far end signal and estimated
// echo using both stored and adapted channels.
// The C code:
// for (i = 0; i < PART_LEN1; i++) {
// echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i],
// far_spectrum[i]);
// (*far_energy) += (uint32_t)(far_spectrum[i]);
// *echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i];
// (*echo_energy_stored) += (uint32_t)echo_est[i];
// }
while (start_stored_p < end_stored_p) {
spectrum_v = vld1q_u16(far_spectrum_p);
adapt_v = vld1q_s16(start_adapt_p);
store_v = vld1q_s16(start_stored_p);
far_energy_v = vaddw_u16(far_energy_v, vget_low_u16(spectrum_v));
far_energy_v = vaddw_u16(far_energy_v, vget_high_u16(spectrum_v));
echo_est_v_low = vmull_u16(vreinterpret_u16_s16(vget_low_s16(store_v)),
vget_low_u16(spectrum_v));
echo_est_v_high = vmull_u16(vreinterpret_u16_s16(vget_high_s16(store_v)),
vget_high_u16(spectrum_v));
vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low));
vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high));
echo_stored_v = vaddq_u32(echo_est_v_low, echo_stored_v);
echo_stored_v = vaddq_u32(echo_est_v_high, echo_stored_v);
echo_adapt_v = vmlal_u16(echo_adapt_v,
vreinterpret_u16_s16(vget_low_s16(adapt_v)),
vget_low_u16(spectrum_v));
echo_adapt_v = vmlal_u16(echo_adapt_v,
vreinterpret_u16_s16(vget_high_s16(adapt_v)),
vget_high_u16(spectrum_v));
start_stored_p += 8;
start_adapt_p += 8;
far_spectrum_p += 8;
echo_est_p += 8;
}
AddLanes(far_energy, far_energy_v);
AddLanes(echo_energy_stored, echo_stored_v);
AddLanes(echo_energy_adapt, echo_adapt_v);
echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN],
far_spectrum[PART_LEN]);
*echo_energy_stored += (uint32_t)echo_est[PART_LEN];
*far_energy += (uint32_t)far_spectrum[PART_LEN];
*echo_energy_adapt += aecm->channelAdapt16[PART_LEN] * far_spectrum[PART_LEN];
}
inline uint16x8_t vld1q(const u16 * ptr) { return vld1q_u16(ptr); }