void test_vshr_nu32 (void) { uint32x2_t out_uint32x2_t; uint32x2_t arg0_uint32x2_t; out_uint32x2_t = vshr_n_u32 (arg0_uint32x2_t, 1); }
int Unpack11to16(const unsigned char* pcInput, unsigned short* pnOutput, const unsigned long nInputSize) { const unsigned char* pOrigInput = pcInput; uint8x8_t inputfield; uint16x4_t shiftfield; uint16_t test[4]; unsigned long nElements = nInputSize / XN_INPUT_ELEMENT_SIZE; // floored unsigned long nNeededOutput = nElements * XN_OUTPUT_ELEMENT_SIZE; // Convert the 11bit packed data into 16bit shorts for (unsigned long 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 #ifdef NEON // Load 64 bits of data inputfield = vld1_u8(pcInput); // Reverse it since the endianess is wrong. inputfield = vrev16_u8(inputfield); // Debug -- let's make sure it looks ok by looking at // it as a 16-bit element since that is ultimately what we want vst1_u16(test, inputfield); printf("i %04x %04x %04x %04x\n", test[0], test[1], test[2], test[3]); // Right shift by 5 bits to aling the first half-word // *note this does not compile since the compiler cannot deal with this // conversion for some reason. It can deal with vshr_n_u32() and lower. // print out the results shiftfield = vshr_n_u64(inputfield, 5); vst1_u16( test,shiftfield); printf("1 %04x %04x %04x %04x\n", test[0], test[1], test[2], test[3]); // Right shift by 10 bits to aling the second half-word // print out the results shiftfield = vshr_n_u32(inputfield, 10); vst1_u16( test,shiftfield); printf("2 %04x %04x %04x %04x\n", test[0], test[1], test[2], test[3]); // Right shift by 15 bits to aling the third half-word // print out the results shiftfield = vshr_n_u32(inputfield, 15); vst1_u16( test,shiftfield); printf("3 %04x %04x %04x %04x\n", test[0], test[1], test[2], test[3]); // we would continue for all 8 half-word results #else // This is the original Primesense code... // shift the output by 5 bits to the right to align 11 bits on the 16 bit field vsri_n_u64(leftfield, shiftfield, 5); vst1_u64((uint64_t*)pnOutput, shiftfield); pnOutput[0] = GetOutput((XN_TAKE_BITS(pcInput[0],8,0) << 3) | XN_TAKE_BITS(pcInput[1],3,5)); pnOutput[1] = GetOutput((XN_TAKE_BITS(pcInput[1],5,0) << 6) | XN_TAKE_BITS(pcInput[2],6,2)); pnOutput[2] = GetOutput((XN_TAKE_BITS(pcInput[2],2,0) << 9) | (XN_TAKE_BITS(pcInput[3],8,0) << 1) | XN_TAKE_BITS(pcInput[4],1,7)); pnOutput[3] = GetOutput((XN_TAKE_BITS(pcInput[4],7,0) << 4) | XN_TAKE_BITS(pcInput[5],4,4)); pnOutput[4] = GetOutput((XN_TAKE_BITS(pcInput[5],4,0) << 7) | XN_TAKE_BITS(pcInput[6],7,1)); pnOutput[5] = GetOutput((XN_TAKE_BITS(pcInput[6],1,0) << 10) | (XN_TAKE_BITS(pcInput[7],8,0) << 2) | XN_TAKE_BITS(pcInput[8],2,6)); pnOutput[6] = GetOutput((XN_TAKE_BITS(pcInput[8],6,0) << 5) | XN_TAKE_BITS(pcInput[9],5,3)); pnOutput[7] = GetOutput((XN_TAKE_BITS(pcInput[9],3,0) << 8) | XN_TAKE_BITS(pcInput[10],8,0)); #endif pcInput += XN_INPUT_ELEMENT_SIZE; pnOutput += 8; } return (pcInput - pOrigInput); }
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 }
unsigned int vp8_variance_halfpixvar16x16_hv_neon( const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int recon_stride, unsigned int *sse) { int i; uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8; int16x4_t d0s16, d1s16, d2s16, d3s16, d10s16, d11s16, d12s16, d13s16; int16x4_t d18s16, d19s16, d20s16, d21s16, d22s16, d23s16, d24s16, d25s16; uint32x2_t d0u32, d10u32; int64x1_t d0s64, d1s64, d2s64, d3s64; uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8; uint16x8_t q0u16, q1u16, q5u16, q6u16, q9u16, q10u16, q11u16, q12u16; int32x4_t q13s32, q14s32, q15s32; int64x2_t q0s64, q1s64, q5s64; q13s32 = vdupq_n_s32(0); q14s32 = vdupq_n_s32(0); q15s32 = vdupq_n_s32(0); q0u8 = vld1q_u8(src_ptr); q1u8 = vld1q_u8(src_ptr + 16); src_ptr += source_stride; q1u8 = vextq_u8(q0u8, q1u8, 1); q0u8 = vrhaddq_u8(q0u8, q1u8); for (i = 0; i < 4; i++) { // vp8_filt_fpo16x16s_4_0_loop_neon q2u8 = vld1q_u8(src_ptr); q3u8 = vld1q_u8(src_ptr + 16); src_ptr += source_stride; q4u8 = vld1q_u8(src_ptr); q5u8 = vld1q_u8(src_ptr + 16); src_ptr += source_stride; q6u8 = vld1q_u8(src_ptr); q7u8 = vld1q_u8(src_ptr + 16); src_ptr += source_stride; q8u8 = vld1q_u8(src_ptr); q9u8 = vld1q_u8(src_ptr + 16); src_ptr += source_stride; q3u8 = vextq_u8(q2u8, q3u8, 1); q5u8 = vextq_u8(q4u8, q5u8, 1); q7u8 = vextq_u8(q6u8, q7u8, 1); q9u8 = vextq_u8(q8u8, q9u8, 1); q1u8 = vrhaddq_u8(q2u8, q3u8); q2u8 = vrhaddq_u8(q4u8, q5u8); q3u8 = vrhaddq_u8(q6u8, q7u8); q4u8 = vrhaddq_u8(q8u8, q9u8); q0u8 = vrhaddq_u8(q0u8, q1u8); q1u8 = vrhaddq_u8(q1u8, q2u8); q2u8 = vrhaddq_u8(q2u8, q3u8); q3u8 = vrhaddq_u8(q3u8, q4u8); q5u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; q6u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; q7u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; q8u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; d0u8 = vget_low_u8(q0u8); d1u8 = vget_high_u8(q0u8); d2u8 = vget_low_u8(q1u8); d3u8 = vget_high_u8(q1u8); d4u8 = vget_low_u8(q2u8); d5u8 = vget_high_u8(q2u8); d6u8 = vget_low_u8(q3u8); d7u8 = vget_high_u8(q3u8); q9u16 = vsubl_u8(d0u8, vget_low_u8(q5u8)); q10u16 = vsubl_u8(d1u8, vget_high_u8(q5u8)); q11u16 = vsubl_u8(d2u8, vget_low_u8(q6u8)); q12u16 = vsubl_u8(d3u8, vget_high_u8(q6u8)); q0u16 = vsubl_u8(d4u8, vget_low_u8(q7u8)); q1u16 = vsubl_u8(d5u8, vget_high_u8(q7u8)); q5u16 = vsubl_u8(d6u8, vget_low_u8(q8u8)); q6u16 = vsubl_u8(d7u8, vget_high_u8(q8u8)); d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16)); d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q9u16)); q14s32 = vmlal_s16(q14s32, d18s16, d18s16); q15s32 = vmlal_s16(q15s32, d19s16, d19s16); d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16)); d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q10u16)); q14s32 = vmlal_s16(q14s32, d20s16, d20s16); q15s32 = vmlal_s16(q15s32, d21s16, d21s16); d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q11u16)); q14s32 = vmlal_s16(q14s32, d22s16, d22s16); q15s32 = vmlal_s16(q15s32, d23s16, d23s16); d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q12u16)); q14s32 = vmlal_s16(q14s32, d24s16, d24s16); q15s32 = vmlal_s16(q15s32, d25s16, d25s16); d0s16 = vreinterpret_s16_u16(vget_low_u16(q0u16)); d1s16 = vreinterpret_s16_u16(vget_high_u16(q0u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q0u16)); q14s32 = vmlal_s16(q14s32, d0s16, d0s16); q15s32 = vmlal_s16(q15s32, d1s16, d1s16); d2s16 = vreinterpret_s16_u16(vget_low_u16(q1u16)); d3s16 = vreinterpret_s16_u16(vget_high_u16(q1u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q1u16)); q14s32 = vmlal_s16(q14s32, d2s16, d2s16); q15s32 = vmlal_s16(q15s32, d3s16, d3s16); d10s16 = vreinterpret_s16_u16(vget_low_u16(q5u16)); d11s16 = vreinterpret_s16_u16(vget_high_u16(q5u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q5u16)); q14s32 = vmlal_s16(q14s32, d10s16, d10s16); q15s32 = vmlal_s16(q15s32, d11s16, d11s16); d12s16 = vreinterpret_s16_u16(vget_low_u16(q6u16)); d13s16 = vreinterpret_s16_u16(vget_high_u16(q6u16)); q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q6u16)); q14s32 = vmlal_s16(q14s32, d12s16, d12s16); q15s32 = vmlal_s16(q15s32, d13s16, d13s16); q0u8 = q4u8; } q15s32 = vaddq_s32(q14s32, q15s32); q0s64 = vpaddlq_s32(q13s32); q1s64 = vpaddlq_s32(q15s32); d0s64 = vget_low_s64(q0s64); d1s64 = vget_high_s64(q0s64); d2s64 = vget_low_s64(q1s64); d3s64 = vget_high_s64(q1s64); d0s64 = vadd_s64(d0s64, d1s64); d1s64 = vadd_s64(d2s64, d3s64); q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64)); vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0); d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8); d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32); return vget_lane_u32(d0u32, 0); }
unsigned int vp8_sub_pixel_variance16x16_neon_func( const unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, const unsigned char *dst_ptr, int dst_pixels_per_line, unsigned int *sse) { int i; DECLARE_ALIGNED_ARRAY(16, unsigned char, tmp, 528); unsigned char *tmpp; unsigned char *tmpp2; uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8; uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d16u8, d17u8, d18u8; uint8x8_t d19u8, d20u8, d21u8; int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16; uint32x2_t d0u32, d10u32; int64x1_t d0s64, d1s64, d2s64, d3s64; uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8; uint8x16_t q10u8, q11u8, q12u8, q13u8, q14u8, q15u8; uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16, q6u16, q7u16, q8u16; uint16x8_t q9u16, q10u16, q11u16, q12u16, q13u16, q14u16; int32x4_t q8s32, q9s32, q10s32; int64x2_t q0s64, q1s64, q5s64; tmpp2 = tmp + 272; tmpp = tmp; if (xoffset == 0) { // secondpass_bfilter16x16_only d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]); d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]); q11u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line; for (i = 4; i > 0; i--) { q12u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line; q13u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line; q14u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line; q15u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line; __builtin_prefetch(src_ptr); __builtin_prefetch(src_ptr + src_pixels_per_line); __builtin_prefetch(src_ptr + src_pixels_per_line * 2); q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8); q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8); q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8); q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8); q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8); q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8); q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8); q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8); q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8); q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8); q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8); q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8); q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8); q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8); q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8); q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8); d2u8 = vqrshrn_n_u16(q1u16, 7); d3u8 = vqrshrn_n_u16(q2u16, 7); d4u8 = vqrshrn_n_u16(q3u16, 7); d5u8 = vqrshrn_n_u16(q4u16, 7); d6u8 = vqrshrn_n_u16(q5u16, 7); d7u8 = vqrshrn_n_u16(q6u16, 7); d8u8 = vqrshrn_n_u16(q7u16, 7); d9u8 = vqrshrn_n_u16(q8u16, 7); q1u8 = vcombine_u8(d2u8, d3u8); q2u8 = vcombine_u8(d4u8, d5u8); q3u8 = vcombine_u8(d6u8, d7u8); q4u8 = vcombine_u8(d8u8, d9u8); q11u8 = q15u8; vst1q_u8((uint8_t *)tmpp2, q1u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q2u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q3u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q4u8); tmpp2 += 16; } } else if (yoffset == 0) { // firstpass_bfilter16x16_only d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]); d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]); for (i = 4; i > 0 ; i--) { d2u8 = vld1_u8(src_ptr); d3u8 = vld1_u8(src_ptr + 8); d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d5u8 = vld1_u8(src_ptr); d6u8 = vld1_u8(src_ptr + 8); d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d8u8 = vld1_u8(src_ptr); d9u8 = vld1_u8(src_ptr + 8); d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d11u8 = vld1_u8(src_ptr); d12u8 = vld1_u8(src_ptr + 8); d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; __builtin_prefetch(src_ptr); __builtin_prefetch(src_ptr + src_pixels_per_line); __builtin_prefetch(src_ptr + src_pixels_per_line * 2); q7u16 = vmull_u8(d2u8, d0u8); q8u16 = vmull_u8(d3u8, d0u8); q9u16 = vmull_u8(d5u8, d0u8); q10u16 = vmull_u8(d6u8, d0u8); q11u16 = vmull_u8(d8u8, d0u8); q12u16 = vmull_u8(d9u8, d0u8); q13u16 = vmull_u8(d11u8, d0u8); q14u16 = vmull_u8(d12u8, d0u8); d2u8 = vext_u8(d2u8, d3u8, 1); d5u8 = vext_u8(d5u8, d6u8, 1); d8u8 = vext_u8(d8u8, d9u8, 1); d11u8 = vext_u8(d11u8, d12u8, 1); q7u16 = vmlal_u8(q7u16, d2u8, d1u8); q9u16 = vmlal_u8(q9u16, d5u8, d1u8); q11u16 = vmlal_u8(q11u16, d8u8, d1u8); q13u16 = vmlal_u8(q13u16, d11u8, d1u8); d3u8 = vext_u8(d3u8, d4u8, 1); d6u8 = vext_u8(d6u8, d7u8, 1); d9u8 = vext_u8(d9u8, d10u8, 1); d12u8 = vext_u8(d12u8, d13u8, 1); q8u16 = vmlal_u8(q8u16, d3u8, d1u8); q10u16 = vmlal_u8(q10u16, d6u8, d1u8); q12u16 = vmlal_u8(q12u16, d9u8, d1u8); q14u16 = vmlal_u8(q14u16, d12u8, d1u8); d14u8 = vqrshrn_n_u16(q7u16, 7); d15u8 = vqrshrn_n_u16(q8u16, 7); d16u8 = vqrshrn_n_u16(q9u16, 7); d17u8 = vqrshrn_n_u16(q10u16, 7); d18u8 = vqrshrn_n_u16(q11u16, 7); d19u8 = vqrshrn_n_u16(q12u16, 7); d20u8 = vqrshrn_n_u16(q13u16, 7); d21u8 = vqrshrn_n_u16(q14u16, 7); q7u8 = vcombine_u8(d14u8, d15u8); q8u8 = vcombine_u8(d16u8, d17u8); q9u8 = vcombine_u8(d18u8, d19u8); q10u8 = vcombine_u8(d20u8, d21u8); vst1q_u8((uint8_t *)tmpp2, q7u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q8u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q9u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q10u8); tmpp2 += 16; } } else { d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]); d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]); d2u8 = vld1_u8(src_ptr); d3u8 = vld1_u8(src_ptr + 8); d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d5u8 = vld1_u8(src_ptr); d6u8 = vld1_u8(src_ptr + 8); d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d8u8 = vld1_u8(src_ptr); d9u8 = vld1_u8(src_ptr + 8); d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d11u8 = vld1_u8(src_ptr); d12u8 = vld1_u8(src_ptr + 8); d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; // First Pass: output_height lines x output_width columns (17x16) for (i = 3; i > 0; i--) { q7u16 = vmull_u8(d2u8, d0u8); q8u16 = vmull_u8(d3u8, d0u8); q9u16 = vmull_u8(d5u8, d0u8); q10u16 = vmull_u8(d6u8, d0u8); q11u16 = vmull_u8(d8u8, d0u8); q12u16 = vmull_u8(d9u8, d0u8); q13u16 = vmull_u8(d11u8, d0u8); q14u16 = vmull_u8(d12u8, d0u8); d2u8 = vext_u8(d2u8, d3u8, 1); d5u8 = vext_u8(d5u8, d6u8, 1); d8u8 = vext_u8(d8u8, d9u8, 1); d11u8 = vext_u8(d11u8, d12u8, 1); q7u16 = vmlal_u8(q7u16, d2u8, d1u8); q9u16 = vmlal_u8(q9u16, d5u8, d1u8); q11u16 = vmlal_u8(q11u16, d8u8, d1u8); q13u16 = vmlal_u8(q13u16, d11u8, d1u8); d3u8 = vext_u8(d3u8, d4u8, 1); d6u8 = vext_u8(d6u8, d7u8, 1); d9u8 = vext_u8(d9u8, d10u8, 1); d12u8 = vext_u8(d12u8, d13u8, 1); q8u16 = vmlal_u8(q8u16, d3u8, d1u8); q10u16 = vmlal_u8(q10u16, d6u8, d1u8); q12u16 = vmlal_u8(q12u16, d9u8, d1u8); q14u16 = vmlal_u8(q14u16, d12u8, d1u8); d14u8 = vqrshrn_n_u16(q7u16, 7); d15u8 = vqrshrn_n_u16(q8u16, 7); d16u8 = vqrshrn_n_u16(q9u16, 7); d17u8 = vqrshrn_n_u16(q10u16, 7); d18u8 = vqrshrn_n_u16(q11u16, 7); d19u8 = vqrshrn_n_u16(q12u16, 7); d20u8 = vqrshrn_n_u16(q13u16, 7); d21u8 = vqrshrn_n_u16(q14u16, 7); d2u8 = vld1_u8(src_ptr); d3u8 = vld1_u8(src_ptr + 8); d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d5u8 = vld1_u8(src_ptr); d6u8 = vld1_u8(src_ptr + 8); d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d8u8 = vld1_u8(src_ptr); d9u8 = vld1_u8(src_ptr + 8); d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; d11u8 = vld1_u8(src_ptr); d12u8 = vld1_u8(src_ptr + 8); d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; q7u8 = vcombine_u8(d14u8, d15u8); q8u8 = vcombine_u8(d16u8, d17u8); q9u8 = vcombine_u8(d18u8, d19u8); q10u8 = vcombine_u8(d20u8, d21u8); vst1q_u8((uint8_t *)tmpp, q7u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q8u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q9u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q10u8); tmpp += 16; } // First-pass filtering for rest 5 lines d14u8 = vld1_u8(src_ptr); d15u8 = vld1_u8(src_ptr + 8); d16u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line; q9u16 = vmull_u8(d2u8, d0u8); q10u16 = vmull_u8(d3u8, d0u8); q11u16 = vmull_u8(d5u8, d0u8); q12u16 = vmull_u8(d6u8, d0u8); q13u16 = vmull_u8(d8u8, d0u8); q14u16 = vmull_u8(d9u8, d0u8); d2u8 = vext_u8(d2u8, d3u8, 1); d5u8 = vext_u8(d5u8, d6u8, 1); d8u8 = vext_u8(d8u8, d9u8, 1); q9u16 = vmlal_u8(q9u16, d2u8, d1u8); q11u16 = vmlal_u8(q11u16, d5u8, d1u8); q13u16 = vmlal_u8(q13u16, d8u8, d1u8); d3u8 = vext_u8(d3u8, d4u8, 1); d6u8 = vext_u8(d6u8, d7u8, 1); d9u8 = vext_u8(d9u8, d10u8, 1); q10u16 = vmlal_u8(q10u16, d3u8, d1u8); q12u16 = vmlal_u8(q12u16, d6u8, d1u8); q14u16 = vmlal_u8(q14u16, d9u8, d1u8); q1u16 = vmull_u8(d11u8, d0u8); q2u16 = vmull_u8(d12u8, d0u8); q3u16 = vmull_u8(d14u8, d0u8); q4u16 = vmull_u8(d15u8, d0u8); d11u8 = vext_u8(d11u8, d12u8, 1); d14u8 = vext_u8(d14u8, d15u8, 1); q1u16 = vmlal_u8(q1u16, d11u8, d1u8); q3u16 = vmlal_u8(q3u16, d14u8, d1u8); d12u8 = vext_u8(d12u8, d13u8, 1); d15u8 = vext_u8(d15u8, d16u8, 1); q2u16 = vmlal_u8(q2u16, d12u8, d1u8); q4u16 = vmlal_u8(q4u16, d15u8, d1u8); d10u8 = vqrshrn_n_u16(q9u16, 7); d11u8 = vqrshrn_n_u16(q10u16, 7); d12u8 = vqrshrn_n_u16(q11u16, 7); d13u8 = vqrshrn_n_u16(q12u16, 7); d14u8 = vqrshrn_n_u16(q13u16, 7); d15u8 = vqrshrn_n_u16(q14u16, 7); d16u8 = vqrshrn_n_u16(q1u16, 7); d17u8 = vqrshrn_n_u16(q2u16, 7); d18u8 = vqrshrn_n_u16(q3u16, 7); d19u8 = vqrshrn_n_u16(q4u16, 7); q5u8 = vcombine_u8(d10u8, d11u8); q6u8 = vcombine_u8(d12u8, d13u8); q7u8 = vcombine_u8(d14u8, d15u8); q8u8 = vcombine_u8(d16u8, d17u8); q9u8 = vcombine_u8(d18u8, d19u8); vst1q_u8((uint8_t *)tmpp, q5u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q6u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q7u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q8u8); tmpp += 16; vst1q_u8((uint8_t *)tmpp, q9u8); // secondpass_filter d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]); d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]); tmpp = tmp; tmpp2 = tmpp + 272; q11u8 = vld1q_u8(tmpp); tmpp += 16; for (i = 4; i > 0; i--) { q12u8 = vld1q_u8(tmpp); tmpp += 16; q13u8 = vld1q_u8(tmpp); tmpp += 16; q14u8 = vld1q_u8(tmpp); tmpp += 16; q15u8 = vld1q_u8(tmpp); tmpp += 16; q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8); q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8); q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8); q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8); q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8); q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8); q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8); q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8); q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8); q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8); q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8); q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8); q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8); q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8); q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8); q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8); d2u8 = vqrshrn_n_u16(q1u16, 7); d3u8 = vqrshrn_n_u16(q2u16, 7); d4u8 = vqrshrn_n_u16(q3u16, 7); d5u8 = vqrshrn_n_u16(q4u16, 7); d6u8 = vqrshrn_n_u16(q5u16, 7); d7u8 = vqrshrn_n_u16(q6u16, 7); d8u8 = vqrshrn_n_u16(q7u16, 7); d9u8 = vqrshrn_n_u16(q8u16, 7); q1u8 = vcombine_u8(d2u8, d3u8); q2u8 = vcombine_u8(d4u8, d5u8); q3u8 = vcombine_u8(d6u8, d7u8); q4u8 = vcombine_u8(d8u8, d9u8); q11u8 = q15u8; vst1q_u8((uint8_t *)tmpp2, q1u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q2u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q3u8); tmpp2 += 16; vst1q_u8((uint8_t *)tmpp2, q4u8); tmpp2 += 16; } } // sub_pixel_variance16x16_neon q8s32 = vdupq_n_s32(0); q9s32 = vdupq_n_s32(0); q10s32 = vdupq_n_s32(0); tmpp = tmp + 272; for (i = 0; i < 8; i++) { // sub_pixel_variance16x16_neon_loop q0u8 = vld1q_u8(tmpp); tmpp += 16; q1u8 = vld1q_u8(tmpp); tmpp += 16; q2u8 = vld1q_u8(dst_ptr); dst_ptr += dst_pixels_per_line; q3u8 = vld1q_u8(dst_ptr); dst_ptr += dst_pixels_per_line; d0u8 = vget_low_u8(q0u8); d1u8 = vget_high_u8(q0u8); d2u8 = vget_low_u8(q1u8); d3u8 = vget_high_u8(q1u8); q11u16 = vsubl_u8(d0u8, vget_low_u8(q2u8)); q12u16 = vsubl_u8(d1u8, vget_high_u8(q2u8)); q13u16 = vsubl_u8(d2u8, vget_low_u8(q3u8)); q14u16 = vsubl_u8(d3u8, vget_high_u8(q3u8)); d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16)); q9s32 = vmlal_s16(q9s32, d22s16, d22s16); q10s32 = vmlal_s16(q10s32, d23s16, d23s16); d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16)); q9s32 = vmlal_s16(q9s32, d24s16, d24s16); q10s32 = vmlal_s16(q10s32, d25s16, d25s16); d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16)); d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16)); q9s32 = vmlal_s16(q9s32, d26s16, d26s16); q10s32 = vmlal_s16(q10s32, d27s16, d27s16); d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16)); d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16)); q9s32 = vmlal_s16(q9s32, d28s16, d28s16); q10s32 = vmlal_s16(q10s32, d29s16, d29s16); } q10s32 = vaddq_s32(q10s32, q9s32); q0s64 = vpaddlq_s32(q8s32); q1s64 = vpaddlq_s32(q10s32); d0s64 = vget_low_s64(q0s64); d1s64 = vget_high_s64(q0s64); d2s64 = vget_low_s64(q1s64); d3s64 = vget_high_s64(q1s64); d0s64 = vadd_s64(d0s64, d1s64); d1s64 = vadd_s64(d2s64, d3s64); q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64)); vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0); d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8); d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32); return vget_lane_u32(d0u32, 0); }
unsigned int vp8_variance16x8_neon( const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int recon_stride, unsigned int *sse) { int i; int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16; uint32x2_t d0u32, d10u32; int64x1_t d0s64, d1s64; uint8x16_t q0u8, q1u8, q2u8, q3u8; uint16x8_t q11u16, q12u16, q13u16, q14u16; int32x4_t q8s32, q9s32, q10s32; int64x2_t q0s64, q1s64, q5s64; q8s32 = vdupq_n_s32(0); q9s32 = vdupq_n_s32(0); q10s32 = vdupq_n_s32(0); for (i = 0; i < 4; i++) { // variance16x8_neon_loop q0u8 = vld1q_u8(src_ptr); src_ptr += source_stride; q1u8 = vld1q_u8(src_ptr); src_ptr += source_stride; __builtin_prefetch(src_ptr); q2u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; q3u8 = vld1q_u8(ref_ptr); ref_ptr += recon_stride; __builtin_prefetch(ref_ptr); q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8)); q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8)); q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8)); q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8)); d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16)); q9s32 = vmlal_s16(q9s32, d22s16, d22s16); q10s32 = vmlal_s16(q10s32, d23s16, d23s16); d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16)); q9s32 = vmlal_s16(q9s32, d24s16, d24s16); q10s32 = vmlal_s16(q10s32, d25s16, d25s16); d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16)); d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16)); q9s32 = vmlal_s16(q9s32, d26s16, d26s16); q10s32 = vmlal_s16(q10s32, d27s16, d27s16); d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16)); d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16)); q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16)); q9s32 = vmlal_s16(q9s32, d28s16, d28s16); q10s32 = vmlal_s16(q10s32, d29s16, d29s16); } q10s32 = vaddq_s32(q10s32, q9s32); q0s64 = vpaddlq_s32(q8s32); q1s64 = vpaddlq_s32(q10s32); d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64)); d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64)); q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64)); vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0); d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7); d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32); return vget_lane_u32(d0u32, 0); }