void AudioBlockPanStereoToStereo_NEON(const float aInputL[WEBAUDIO_BLOCK_SIZE],
const float aInputR[WEBAUDIO_BLOCK_SIZE],
float aGainL, float aGainR,
bool aIsOnTheLeft,
float aOutputL[WEBAUDIO_BLOCK_SIZE],
float aOutputR[WEBAUDIO_BLOCK_SIZE]) {
ASSERT_ALIGNED(aInputL);
ASSERT_ALIGNED(aInputR);
ASSERT_ALIGNED(aOutputL);
ASSERT_ALIGNED(aOutputR);
float32x4_t vinL0, vinL1;
float32x4_t vinR0, vinR1;
float32x4_t voutL0, voutL1;
float32x4_t voutR0, voutR1;
float32x4_t vscaleL = vmovq_n_f32(aGainL);
float32x4_t vscaleR = vmovq_n_f32(aGainR);
if (aIsOnTheLeft) {
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
voutL0 = vmlaq_f32(vinL0, vinR0, vscaleL);
voutL1 = vmlaq_f32(vinL1, vinR1, vscaleL);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
voutR0 = vmulq_f32(vinR0, vscaleR);
voutR1 = vmulq_f32(vinR1, vscaleR);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
} else {
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
voutL0 = vmulq_f32(vinL0, vscaleL);
voutL1 = vmulq_f32(vinL1, vscaleL);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
voutR0 = vmlaq_f32(vinR0, vinL0, vscaleR);
voutR1 = vmlaq_f32(vinR1, vinL1, vscaleR);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
}
}
void
AudioBlockCopyChannelWithScale_NEON(const float aInput[WEBAUDIO_BLOCK_SIZE],
const float aScale[WEBAUDIO_BLOCK_SIZE],
float aOutput[WEBAUDIO_BLOCK_SIZE])
{
ASSERT_ALIGNED(aInput);
ASSERT_ALIGNED(aScale);
ASSERT_ALIGNED(aOutput);
float32x4_t vin0, vin1, vin2, vin3;
float32x4_t vout0, vout1, vout2, vout3;
float32x4_t vscale0, vscale1, vscale2, vscale3;
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i+=16) {
vin0 = vld1q_f32(ADDRESS_OF(aInput, i));
vin1 = vld1q_f32(ADDRESS_OF(aInput, i+4));
vin2 = vld1q_f32(ADDRESS_OF(aInput, i+8));
vin3 = vld1q_f32(ADDRESS_OF(aInput, i+12));
vscale0 = vld1q_f32(ADDRESS_OF(aScale, i));
vscale1 = vld1q_f32(ADDRESS_OF(aScale, i+4));
vscale2 = vld1q_f32(ADDRESS_OF(aScale, i+8));
vscale3 = vld1q_f32(ADDRESS_OF(aScale, i+12));
vout0 = vmulq_f32(vin0, vscale0);
vout1 = vmulq_f32(vin1, vscale1);
vout2 = vmulq_f32(vin2, vscale2);
vout3 = vmulq_f32(vin3, vscale3);
vst1q_f32(ADDRESS_OF(aOutput, i), vout0);
vst1q_f32(ADDRESS_OF(aOutput, i+4), vout1);
vst1q_f32(ADDRESS_OF(aOutput, i+8), vout2);
vst1q_f32(ADDRESS_OF(aOutput, i+12), vout3);
}
}
void AudioBufferInPlaceScale_NEON(float* aBlock, float aScale, uint32_t aSize) {
ASSERT_ALIGNED(aBlock);
float32x4_t vin0, vin1, vin2, vin3;
float32x4_t vout0, vout1, vout2, vout3;
float32x4_t vscale = vmovq_n_f32(aScale);
uint32_t dif = aSize % 16;
uint32_t vectorSize = aSize - dif;
uint32_t i = 0;
for (; i < vectorSize; i += 16) {
vin0 = vld1q_f32(ADDRESS_OF(aBlock, i));
vin1 = vld1q_f32(ADDRESS_OF(aBlock, i + 4));
vin2 = vld1q_f32(ADDRESS_OF(aBlock, i + 8));
vin3 = vld1q_f32(ADDRESS_OF(aBlock, i + 12));
vout0 = vmulq_f32(vin0, vscale);
vout1 = vmulq_f32(vin1, vscale);
vout2 = vmulq_f32(vin2, vscale);
vout3 = vmulq_f32(vin3, vscale);
vst1q_f32(ADDRESS_OF(aBlock, i), vout0);
vst1q_f32(ADDRESS_OF(aBlock, i + 4), vout1);
vst1q_f32(ADDRESS_OF(aBlock, i + 8), vout2);
vst1q_f32(ADDRESS_OF(aBlock, i + 12), vout3);
}
for (unsigned j = 0; j < dif; ++i, ++j) {
aBlock[i] *= aScale;
}
}
void test_alignment (int max_size = 2000)
{
std::vector<V *, nonstd::aligned_allocator<V *>> vectors(0);
for (int size = 0; size < max_size; ++size) {
vectors.push_back(new V(size));
}
for (int size = 0; size < max_size; ++size) {
V * p = vectors[size];
ASSERT_ALIGNED(& (* p)[0]);
delete p;
}
ASSERT_ALIGNED(& vectors[0]);
vectors.clear();
}
void AudioBufferAddWithScale_NEON(const float* aInput,
float aScale,
float* aOutput,
uint32_t aSize)
{
ASSERT_ALIGNED(aInput);
ASSERT_ALIGNED(aOutput);
float32x4_t vin0, vin1, vin2, vin3;
float32x4_t vout0, vout1, vout2, vout3;
float32x4_t vscale = vmovq_n_f32(aScale);
uint32_t dif = aSize % 16;
aSize -= dif;
unsigned i = 0;
for (; i < aSize; i+=16) {
vin0 = vld1q_f32(ADDRESS_OF(aInput, i));
vin1 = vld1q_f32(ADDRESS_OF(aInput, i+4));
vin2 = vld1q_f32(ADDRESS_OF(aInput, i+8));
vin3 = vld1q_f32(ADDRESS_OF(aInput, i+12));
vout0 = vld1q_f32(ADDRESS_OF(aOutput, i));
vout1 = vld1q_f32(ADDRESS_OF(aOutput, i+4));
vout2 = vld1q_f32(ADDRESS_OF(aOutput, i+8));
vout3 = vld1q_f32(ADDRESS_OF(aOutput, i+12));
vout0 = vmlaq_f32(vout0, vin0, vscale);
vout1 = vmlaq_f32(vout1, vin1, vscale);
vout2 = vmlaq_f32(vout2, vin2, vscale);
vout3 = vmlaq_f32(vout3, vin3, vscale);
vst1q_f32(ADDRESS_OF(aOutput, i), vout0);
vst1q_f32(ADDRESS_OF(aOutput, i+4), vout1);
vst1q_f32(ADDRESS_OF(aOutput, i+8), vout2);
vst1q_f32(ADDRESS_OF(aOutput, i+12), vout3);
}
for (unsigned j = 0; j < dif; ++i, ++j) {
aOutput[i] += aInput[i]*aScale;
}
}
/**
* \brief initializes the hardware specific part of the descriptor to be used
* for memcpy descriptors
*
* \param desc Xeon Phi descriptor
* \param src Source address of the transfer
* \param dst destination address of the transfer
* \param size number of bytes to copy
* \param flags control flags
*
* XXX: this function assumes that the size of the descriptor has already been
* checked and must match the maximum transfer size of the channel
*/
inline void xeon_phi_dma_desc_fill_memcpy(struct dma_descriptor *desc,
lpaddr_t src,
lpaddr_t dst,
uint32_t size,
uint32_t flags)
{
uint8_t *d = dma_desc_get_desc_handle(desc);
clear_descriptor(d);
ASSERT_ALIGNED(src);
ASSERT_ALIGNED(dst);
ASSERT_ALIGNED(size);
if (flags & XEON_PHI_DMA_DESC_FLAG_INTR) {
xeon_phi_dma_desc_memcpy_intr_insert(d, 0x1);
}
if (flags & XEON_PHI_DMA_DESC_FLAG_TWB) {
xeon_phi_dma_desc_memcpy_twb_insert(d, 0x1);
}
if (flags & XEON_PHI_DMA_DESC_FLAG_C) {
xeon_phi_dma_desc_memcpy_c_insert(d, 0x1);
}
if (flags & XEON_PHI_DMA_DESC_FLAG_CO) {
xeon_phi_dma_desc_memcpy_co_insert(d, 0x1);
}
if (flags & XEON_PHI_DMA_DESC_FLAG_ECY) {
xeon_phi_dma_desc_memcpy_ecy_insert(d, 0x1);
}
xeon_phi_dma_desc_memcpy_intr_insert(d, 0x1);
xeon_phi_dma_desc_memcpy_twb_insert(d, 0x1);
xeon_phi_dma_desc_memcpy_src_insert(d, src);
xeon_phi_dma_desc_memcpy_dst_insert(d, dst);
xeon_phi_dma_desc_memcpy_length_insert(d, (size >> XEON_PHI_DMA_ALIGN_SHIFT));
xeon_phi_dma_desc_memcpy_dtype_insert(d, xeon_phi_dma_desc_memcpy);
}
/**
* \brief initializes the hardware specific part of the descriptor to be used
* for general descriptors
*
* \param desc Xeon Phi descriptor
* \param dst destination address
* \param data Data payload for the request (request specific)
*/
inline void xeon_phi_dma_desc_fill_general(struct dma_descriptor *desc,
lpaddr_t dst,
uint64_t data)
{
uint8_t *d = dma_desc_get_desc_handle(desc);
clear_descriptor(d);
ASSERT_ALIGNED(dst);
xeon_phi_dma_desc_general_data_insert(d, data);
xeon_phi_dma_desc_general_dst_insert(d, dst);
xeon_phi_dma_desc_general_dtype_insert(d,
xeon_phi_dma_desc_general);
}
/**
* \brief initializes the hardware specific part of the descriptor to be used
* for status descriptors
*
* \param desc Xeon Phi descriptor
* \param dst destination address
* \param data Data payload for the request (request specific)
* \param flags Descriptor flags
*/
inline void xeon_phi_dma_desc_fill_status(struct dma_descriptor *desc,
lpaddr_t dst,
uint64_t data,
uint32_t flags)
{
uint8_t *d = dma_desc_get_desc_handle(desc);
clear_descriptor(d);
ASSERT_ALIGNED(dst);
xeon_phi_dma_desc_status_data_insert(d, data);
xeon_phi_dma_desc_status_dst_insert(d, dst);
if (flags & XEON_PHI_DMA_DESC_FLAG_INTR) {
xeon_phi_dma_desc_status_intr_insert(d, 0x1);
}
xeon_phi_dma_desc_status_dtype_insert(d,
xeon_phi_dma_desc_status);
}
static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 permM2 = vec_lvsl(-2, src);
const vec_u8 permM1 = vec_lvsl(-1, src);
const vec_u8 permP0 = vec_lvsl(+0, src);
const vec_u8 permP1 = vec_lvsl(+1, src);
const vec_u8 permP2 = vec_lvsl(+2, src);
const vec_u8 permP3 = vec_lvsl(+3, src);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 srcP0A, srcP0B, srcP1A, srcP1B,
srcP2A, srcP2B, srcP3A, srcP3B,
srcM1A, srcM1B, srcM2A, srcM2B,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB;
vec_u8 sum, fsum;
for (i = 0 ; i < 16 ; i ++) {
vec_u8 srcR1 = vec_ld(-2, src);
vec_u8 srcR2 = vec_ld(14, src);
switch (align) {
default: {
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = vec_perm(srcR1, srcR2, permP2);
srcP3 = vec_perm(srcR1, srcR2, permP3);
} break;
case 11: {
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = vec_perm(srcR1, srcR2, permP2);
srcP3 = srcR2;
} break;
case 12: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = srcR2;
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 13: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = srcR2;
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 14: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = srcR2;
srcP1 = vec_perm(srcR2, srcR3, permP1);
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 15: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = srcR2;
srcP0 = vec_perm(srcR2, srcR3, permP0);
srcP1 = vec_perm(srcR2, srcR3, permP1);
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
}
srcP0A = (vec_s16) vec_mergeh(zero_u8v, srcP0);
srcP0B = (vec_s16) vec_mergel(zero_u8v, srcP0);
srcP1A = (vec_s16) vec_mergeh(zero_u8v, srcP1);
srcP1B = (vec_s16) vec_mergel(zero_u8v, srcP1);
srcP2A = (vec_s16) vec_mergeh(zero_u8v, srcP2);
srcP2B = (vec_s16) vec_mergel(zero_u8v, srcP2);
srcP3A = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3B = (vec_s16) vec_mergel(zero_u8v, srcP3);
srcM1A = (vec_s16) vec_mergeh(zero_u8v, srcM1);
srcM1B = (vec_s16) vec_mergel(zero_u8v, srcM1);
srcM2A = (vec_s16) vec_mergeh(zero_u8v, srcM2);
srcM2B = (vec_s16) vec_mergel(zero_u8v, srcM2);
sum1A = vec_adds(srcP0A, srcP1A);
sum1B = vec_adds(srcP0B, srcP1B);
sum2A = vec_adds(srcM1A, srcP2A);
sum2B = vec_adds(srcM1B, srcP2B);
sum3A = vec_adds(srcM2A, srcP3A);
sum3B = vec_adds(srcM2B, srcP3B);
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
src += srcStride;
dst += dstStride;
}
}
static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, uint8_t * src, int dstStride, int tmpStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 permM2 = vec_lvsl(-2, src);
const vec_u8 permM1 = vec_lvsl(-1, src);
const vec_u8 permP0 = vec_lvsl(+0, src);
const vec_u8 permP1 = vec_lvsl(+1, src);
const vec_u8 permP2 = vec_lvsl(+2, src);
const vec_u8 permP3 = vec_lvsl(+3, src);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u32 v10ui = vec_splat_u32(10);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v1ss = vec_splat_s16(1);
const vec_s32 v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
const vec_u32 v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 srcP0A, srcP0B, srcP1A, srcP1B,
srcP2A, srcP2B, srcP3A, srcP3B,
srcM1A, srcM1B, srcM2A, srcM2B,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
pp1A, pp1B, pp2A, pp2B, psumA, psumB;
const vec_u8 mperm = (const vec_u8)
{0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F};
int16_t *tmpbis = tmp;
vec_s16 tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB,
tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB,
tmpP2ssA, tmpP2ssB;
vec_s32 pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
ssumAe, ssumAo, ssumBe, ssumBo;
vec_u8 fsum, sumv, sum;
vec_s16 ssume, ssumo;
src -= (2 * srcStride);
for (i = 0 ; i < 21 ; i ++) {
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
vec_u8 srcR1 = vec_ld(-2, src);
vec_u8 srcR2 = vec_ld(14, src);
switch (align) {
default: {
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = vec_perm(srcR1, srcR2, permP2);
srcP3 = vec_perm(srcR1, srcR2, permP3);
} break;
case 11: {
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = vec_perm(srcR1, srcR2, permP2);
srcP3 = srcR2;
} break;
case 12: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = vec_perm(srcR1, srcR2, permP1);
srcP2 = srcR2;
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 13: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = vec_perm(srcR1, srcR2, permP0);
srcP1 = srcR2;
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 14: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = vec_perm(srcR1, srcR2, permM1);
srcP0 = srcR2;
srcP1 = vec_perm(srcR2, srcR3, permP1);
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
case 15: {
vec_u8 srcR3 = vec_ld(30, src);
srcM2 = vec_perm(srcR1, srcR2, permM2);
srcM1 = srcR2;
srcP0 = vec_perm(srcR2, srcR3, permP0);
srcP1 = vec_perm(srcR2, srcR3, permP1);
srcP2 = vec_perm(srcR2, srcR3, permP2);
srcP3 = vec_perm(srcR2, srcR3, permP3);
} break;
}
srcP0A = (vec_s16) vec_mergeh(zero_u8v, srcP0);
srcP0B = (vec_s16) vec_mergel(zero_u8v, srcP0);
srcP1A = (vec_s16) vec_mergeh(zero_u8v, srcP1);
srcP1B = (vec_s16) vec_mergel(zero_u8v, srcP1);
srcP2A = (vec_s16) vec_mergeh(zero_u8v, srcP2);
srcP2B = (vec_s16) vec_mergel(zero_u8v, srcP2);
srcP3A = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3B = (vec_s16) vec_mergel(zero_u8v, srcP3);
srcM1A = (vec_s16) vec_mergeh(zero_u8v, srcM1);
srcM1B = (vec_s16) vec_mergel(zero_u8v, srcM1);
srcM2A = (vec_s16) vec_mergeh(zero_u8v, srcM2);
srcM2B = (vec_s16) vec_mergel(zero_u8v, srcM2);
sum1A = vec_adds(srcP0A, srcP1A);
sum1B = vec_adds(srcP0B, srcP1B);
sum2A = vec_adds(srcM1A, srcP2A);
sum2B = vec_adds(srcM1B, srcP2B);
sum3A = vec_adds(srcM2A, srcP3A);
sum3B = vec_adds(srcM2B, srcP3B);
pp1A = vec_mladd(sum1A, v20ss, sum3A);
pp1B = vec_mladd(sum1B, v20ss, sum3B);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
psumA = vec_sub(pp1A, pp2A);
psumB = vec_sub(pp1B, pp2B);
vec_st(psumA, 0, tmp);
vec_st(psumB, 16, tmp);
src += srcStride;
tmp += tmpStride; /* int16_t*, and stride is 16, so it's OK here */
}
tmpM2ssA = vec_ld(0, tmpbis);
tmpM2ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpM1ssA = vec_ld(0, tmpbis);
tmpM1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP0ssA = vec_ld(0, tmpbis);
tmpP0ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP1ssA = vec_ld(0, tmpbis);
tmpP1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP2ssA = vec_ld(0, tmpbis);
tmpP2ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
for (i = 0 ; i < 16 ; i++) {
const vec_s16 tmpP3ssA = vec_ld(0, tmpbis);
const vec_s16 tmpP3ssB = vec_ld(16, tmpbis);
const vec_s16 sum1A = vec_adds(tmpP0ssA, tmpP1ssA);
const vec_s16 sum1B = vec_adds(tmpP0ssB, tmpP1ssB);
const vec_s16 sum2A = vec_adds(tmpM1ssA, tmpP2ssA);
const vec_s16 sum2B = vec_adds(tmpM1ssB, tmpP2ssB);
const vec_s16 sum3A = vec_adds(tmpM2ssA, tmpP3ssA);
const vec_s16 sum3B = vec_adds(tmpM2ssB, tmpP3ssB);
tmpbis += tmpStride;
tmpM2ssA = tmpM1ssA;
tmpM2ssB = tmpM1ssB;
tmpM1ssA = tmpP0ssA;
tmpM1ssB = tmpP0ssB;
tmpP0ssA = tmpP1ssA;
tmpP0ssB = tmpP1ssB;
tmpP1ssA = tmpP2ssA;
tmpP1ssB = tmpP2ssB;
tmpP2ssA = tmpP3ssA;
tmpP2ssB = tmpP3ssB;
pp1Ae = vec_mule(sum1A, v20ss);
pp1Ao = vec_mulo(sum1A, v20ss);
pp1Be = vec_mule(sum1B, v20ss);
pp1Bo = vec_mulo(sum1B, v20ss);
pp2Ae = vec_mule(sum2A, v5ss);
pp2Ao = vec_mulo(sum2A, v5ss);
pp2Be = vec_mule(sum2B, v5ss);
pp2Bo = vec_mulo(sum2B, v5ss);
pp3Ae = vec_sra((vec_s32)sum3A, v16ui);
pp3Ao = vec_mulo(sum3A, v1ss);
pp3Be = vec_sra((vec_s32)sum3B, v16ui);
pp3Bo = vec_mulo(sum3B, v1ss);
pp1cAe = vec_add(pp1Ae, v512si);
pp1cAo = vec_add(pp1Ao, v512si);
pp1cBe = vec_add(pp1Be, v512si);
pp1cBo = vec_add(pp1Bo, v512si);
pp32Ae = vec_sub(pp3Ae, pp2Ae);
pp32Ao = vec_sub(pp3Ao, pp2Ao);
pp32Be = vec_sub(pp3Be, pp2Be);
pp32Bo = vec_sub(pp3Bo, pp2Bo);
sumAe = vec_add(pp1cAe, pp32Ae);
sumAo = vec_add(pp1cAo, pp32Ao);
sumBe = vec_add(pp1cBe, pp32Be);
sumBo = vec_add(pp1cBo, pp32Bo);
ssumAe = vec_sra(sumAe, v10ui);
ssumAo = vec_sra(sumAo, v10ui);
ssumBe = vec_sra(sumBe, v10ui);
ssumBo = vec_sra(sumBo, v10ui);
ssume = vec_packs(ssumAe, ssumBe);
ssumo = vec_packs(ssumAo, ssumBo);
sumv = vec_packsu(ssume, ssumo);
sum = vec_perm(sumv, sumv, mperm);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 perm = vec_lvsl(0, src);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
uint8_t *srcbis = src - (srcStride * 2);
const vec_u8 srcM2a = vec_ld(0, srcbis);
const vec_u8 srcM2b = vec_ld(16, srcbis);
const vec_u8 srcM2 = vec_perm(srcM2a, srcM2b, perm);
//srcbis += srcStride;
const vec_u8 srcM1a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcM1b = vec_ld(16, srcbis);
const vec_u8 srcM1 = vec_perm(srcM1a, srcM1b, perm);
//srcbis += srcStride;
const vec_u8 srcP0a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP0b = vec_ld(16, srcbis);
const vec_u8 srcP0 = vec_perm(srcP0a, srcP0b, perm);
//srcbis += srcStride;
const vec_u8 srcP1a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP1b = vec_ld(16, srcbis);
const vec_u8 srcP1 = vec_perm(srcP1a, srcP1b, perm);
//srcbis += srcStride;
const vec_u8 srcP2a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP2b = vec_ld(16, srcbis);
const vec_u8 srcP2 = vec_perm(srcP2a, srcP2b, perm);
//srcbis += srcStride;
vec_s16 srcM2ssA = (vec_s16) vec_mergeh(zero_u8v, srcM2);
vec_s16 srcM2ssB = (vec_s16) vec_mergel(zero_u8v, srcM2);
vec_s16 srcM1ssA = (vec_s16) vec_mergeh(zero_u8v, srcM1);
vec_s16 srcM1ssB = (vec_s16) vec_mergel(zero_u8v, srcM1);
vec_s16 srcP0ssA = (vec_s16) vec_mergeh(zero_u8v, srcP0);
vec_s16 srcP0ssB = (vec_s16) vec_mergel(zero_u8v, srcP0);
vec_s16 srcP1ssA = (vec_s16) vec_mergeh(zero_u8v, srcP1);
vec_s16 srcP1ssB = (vec_s16) vec_mergel(zero_u8v, srcP1);
vec_s16 srcP2ssA = (vec_s16) vec_mergeh(zero_u8v, srcP2);
vec_s16 srcP2ssB = (vec_s16) vec_mergel(zero_u8v, srcP2);
vec_s16 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB,
srcP3ssA, srcP3ssB,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
vec_u8 sum, fsum, srcP3a, srcP3b, srcP3;
for (i = 0 ; i < 16 ; i++) {
srcP3a = vec_ld(0, srcbis += srcStride);
srcP3b = vec_ld(16, srcbis);
srcP3 = vec_perm(srcP3a, srcP3b, perm);
srcP3ssA = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3ssB = (vec_s16) vec_mergel(zero_u8v, srcP3);
//srcbis += srcStride;
sum1A = vec_adds(srcP0ssA, srcP1ssA);
sum1B = vec_adds(srcP0ssB, srcP1ssB);
sum2A = vec_adds(srcM1ssA, srcP2ssA);
sum2B = vec_adds(srcM1ssB, srcP2ssB);
sum3A = vec_adds(srcM2ssA, srcP3ssA);
sum3B = vec_adds(srcM2ssB, srcP3ssB);
srcM2ssA = srcM1ssA;
srcM2ssB = srcM1ssB;
srcM1ssA = srcP0ssA;
srcM1ssB = srcP0ssB;
srcP0ssA = srcP1ssA;
srcP0ssB = srcP1ssB;
srcP1ssA = srcP2ssA;
srcP1ssB = srcP2ssB;
srcP2ssA = srcP3ssA;
srcP2ssB = srcP3ssB;
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
Camera::Camera( float w, float h, float near, float far ) {
ASSERT_ALIGNED(this);
setFrustum(w, h, near, far);
}
void AudioBlockPanStereoToStereo_NEON(
const float aInputL[WEBAUDIO_BLOCK_SIZE],
const float aInputR[WEBAUDIO_BLOCK_SIZE], float aGainL[WEBAUDIO_BLOCK_SIZE],
float aGainR[WEBAUDIO_BLOCK_SIZE],
const bool aIsOnTheLeft[WEBAUDIO_BLOCK_SIZE],
float aOutputL[WEBAUDIO_BLOCK_SIZE], float aOutputR[WEBAUDIO_BLOCK_SIZE]) {
ASSERT_ALIGNED(aInputL);
ASSERT_ALIGNED(aInputR);
ASSERT_ALIGNED(aGainL);
ASSERT_ALIGNED(aGainR);
ASSERT_ALIGNED(aIsOnTheLeft);
ASSERT_ALIGNED(aOutputL);
ASSERT_ALIGNED(aOutputR);
float32x4_t vinL0, vinL1;
float32x4_t vinR0, vinR1;
float32x4_t voutL0, voutL1;
float32x4_t voutR0, voutR1;
float32x4_t vscaleL0, vscaleL1;
float32x4_t vscaleR0, vscaleR1;
float32x4_t onleft0, onleft1, notonleft0, notonleft1;
float32x4_t zero = vmovq_n_f32(0);
uint8x8_t isOnTheLeft;
// Although MSVC throws uninitialized value warning for voutL0 and voutL1,
// since we fill all lanes by vsetq_lane_f32, we can ignore it. But to avoid
// compiler warning, set zero.
voutL0 = zero;
voutL1 = zero;
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
vscaleL0 = vld1q_f32(ADDRESS_OF(aGainL, i));
vscaleL1 = vld1q_f32(ADDRESS_OF(aGainL, i + 4));
vscaleR0 = vld1q_f32(ADDRESS_OF(aGainR, i));
vscaleR1 = vld1q_f32(ADDRESS_OF(aGainR, i + 4));
// Load output with boolean "on the left" values. This assumes that
// bools are stored as a single byte.
isOnTheLeft = vld1_u8((uint8_t*)&aIsOnTheLeft[i]);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 0), voutL0, 0);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 1), voutL0, 1);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 2), voutL0, 2);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 3), voutL0, 3);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 4), voutL1, 0);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 5), voutL1, 1);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 6), voutL1, 2);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 7), voutL1, 3);
// Convert the boolean values into masks by setting all bits to 1
// if true.
voutL0 = (float32x4_t)vcgtq_f32(voutL0, zero);
voutL1 = (float32x4_t)vcgtq_f32(voutL1, zero);
// The right output masks are the same as the left masks
voutR0 = voutL0;
voutR1 = voutL1;
// Calculate left channel assuming isOnTheLeft
onleft0 = vmlaq_f32(vinL0, vinR0, vscaleL0);
onleft1 = vmlaq_f32(vinL1, vinR1, vscaleL0);
// Calculate left channel assuming not isOnTheLeft
notonleft0 = vmulq_f32(vinL0, vscaleL0);
notonleft1 = vmulq_f32(vinL1, vscaleL1);
// Write results using previously stored masks
voutL0 = vbslq_f32((uint32x4_t)voutL0, onleft0, notonleft0);
voutL1 = vbslq_f32((uint32x4_t)voutL1, onleft1, notonleft1);
// Calculate right channel assuming isOnTheLeft
onleft0 = vmulq_f32(vinR0, vscaleR0);
onleft1 = vmulq_f32(vinR1, vscaleR1);
// Calculate right channel assuming not isOnTheLeft
notonleft0 = vmlaq_f32(vinR0, vinL0, vscaleR0);
notonleft1 = vmlaq_f32(vinR1, vinL1, vscaleR1);
// Write results using previously stored masks
voutR0 = vbslq_f32((uint32x4_t)voutR0, onleft0, notonleft0);
voutR1 = vbslq_f32((uint32x4_t)voutR1, onleft1, notonleft1);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
}
static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 perm;
#if HAVE_BIGENDIAN
perm = vec_lvsl(0, src);
#endif
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
const uint8_t *srcbis = src - (srcStride * 2);
const vec_u8 srcM2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcM1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP0 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
vec_s16 srcM2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
vec_s16 srcM2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
vec_s16 srcM1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
vec_s16 srcM1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
vec_s16 srcP0ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
vec_s16 srcP0ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
vec_s16 srcP1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
vec_s16 srcP1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
vec_s16 srcP2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
vec_s16 srcP2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
vec_s16 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB,
srcP3ssA, srcP3ssB,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
vec_u8 sum, fsum, srcP3;
for (i = 0 ; i < 16 ; i++) {
srcP3 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
srcP3ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
sum1A = vec_adds(srcP0ssA, srcP1ssA);
sum1B = vec_adds(srcP0ssB, srcP1ssB);
sum2A = vec_adds(srcM1ssA, srcP2ssA);
sum2B = vec_adds(srcM1ssB, srcP2ssB);
sum3A = vec_adds(srcM2ssA, srcP3ssA);
sum3B = vec_adds(srcM2ssB, srcP3ssB);
srcM2ssA = srcM1ssA;
srcM2ssB = srcM1ssB;
srcM1ssA = srcP0ssA;
srcM1ssB = srcP0ssB;
srcP0ssA = srcP1ssA;
srcP0ssB = srcP1ssB;
srcP1ssA = srcP2ssA;
srcP1ssB = srcP2ssB;
srcP2ssA = srcP3ssA;
srcP2ssB = srcP3ssB;
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 permM2, permM1, permP0, permP1, permP2, permP3;
const vec_s16 v5ss = vec_splat_s16(5);
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 srcP0A, srcP0B, srcP1A, srcP1B,
srcP2A, srcP2B, srcP3A, srcP3B,
srcM1A, srcM1B, srcM2A, srcM2B,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB;
vec_u8 sum, fsum;
#if HAVE_BIGENDIAN
permM2 = vec_lvsl(-2, src);
permM1 = vec_lvsl(-1, src);
permP0 = vec_lvsl(+0, src);
permP1 = vec_lvsl(+1, src);
permP2 = vec_lvsl(+2, src);
permP3 = vec_lvsl(+3, src);
#endif /* HAVE_BIGENDIAN */
for (i = 0 ; i < 16 ; i ++) {
load_alignment(src, align, permM2, permM1, permP0, permP1, permP2, permP3);
srcP0A = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
srcP0B = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
srcP1A = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
srcP1B = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
srcP2A = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
srcP2B = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
srcP3A = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3B = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
srcM1A = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
srcM1B = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
srcM2A = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
srcM2B = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
sum1A = vec_adds(srcP0A, srcP1A);
sum1B = vec_adds(srcP0B, srcP1B);
sum2A = vec_adds(srcM1A, srcP2A);
sum2B = vec_adds(srcM1B, srcP2B);
sum3A = vec_adds(srcM2A, srcP3A);
sum3B = vec_adds(srcM2B, srcP3B);
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
src += srcStride;
dst += dstStride;
}
}
inline pack<REAL> aligned_load( ITER aligned_ptr )
{
ASSERT_ALIGNED( aligned_ptr );
return SIMD_INTRIN( load_ps )( &(*aligned_ptr) );
}
Camera::Camera() {
ASSERT_ALIGNED(this);
}
inline void pack<float>::aligned_store( float* mem ) const
{
ASSERT_ALIGNED( mem );
return SIMD_INTRIN( store_ps )( mem, xmm_ );
}
inline void pack<double>::aligned_store( double* mem ) const
{
ASSERT_ALIGNED( mem );
return SIMD_INTRIN( store_pd )( mem, xmmd_ );
}
