diff options
Diffstat (limited to 'third_party/aom/av1/common/arm/transpose_neon.h')
| -rw-r--r-- | third_party/aom/av1/common/arm/transpose_neon.h | 83 |
1 files changed, 80 insertions, 3 deletions
diff --git a/third_party/aom/av1/common/arm/transpose_neon.h b/third_party/aom/av1/common/arm/transpose_neon.h index fe134087b..8a3d9f07f 100644 --- a/third_party/aom/av1/common/arm/transpose_neon.h +++ b/third_party/aom/av1/common/arm/transpose_neon.h @@ -8,8 +8,8 @@ * be found in the AUTHORS file in the root of the source tree. */ -#ifndef AV1_COMMON_ARM_TRANSPOSE_NEON_H_ -#define AV1_COMMON_ARM_TRANSPOSE_NEON_H_ +#ifndef AOM_AV1_COMMON_ARM_TRANSPOSE_NEON_H_ +#define AOM_AV1_COMMON_ARM_TRANSPOSE_NEON_H_ #include <arm_neon.h> @@ -386,6 +386,83 @@ static INLINE void transpose_s16_8x8(int16x8_t *a0, int16x8_t *a1, vget_high_s16(vreinterpretq_s16_s32(c3.val[1]))); } +static INLINE int16x8x2_t vpx_vtrnq_s64_to_s16(int32x4_t a0, int32x4_t a1) { + int16x8x2_t b0; + b0.val[0] = vcombine_s16(vreinterpret_s16_s32(vget_low_s32(a0)), + vreinterpret_s16_s32(vget_low_s32(a1))); + b0.val[1] = vcombine_s16(vreinterpret_s16_s32(vget_high_s32(a0)), + vreinterpret_s16_s32(vget_high_s32(a1))); + return b0; +} + +static INLINE void transpose_s16_8x8q(int16x8_t *a0, int16x8_t *out) { + // Swap 16 bit elements. Goes from: + // a0: 00 01 02 03 04 05 06 07 + // a1: 10 11 12 13 14 15 16 17 + // a2: 20 21 22 23 24 25 26 27 + // a3: 30 31 32 33 34 35 36 37 + // a4: 40 41 42 43 44 45 46 47 + // a5: 50 51 52 53 54 55 56 57 + // a6: 60 61 62 63 64 65 66 67 + // a7: 70 71 72 73 74 75 76 77 + // to: + // b0.val[0]: 00 10 02 12 04 14 06 16 + // b0.val[1]: 01 11 03 13 05 15 07 17 + // b1.val[0]: 20 30 22 32 24 34 26 36 + // b1.val[1]: 21 31 23 33 25 35 27 37 + // b2.val[0]: 40 50 42 52 44 54 46 56 + // b2.val[1]: 41 51 43 53 45 55 47 57 + // b3.val[0]: 60 70 62 72 64 74 66 76 + // b3.val[1]: 61 71 63 73 65 75 67 77 + + const int16x8x2_t b0 = vtrnq_s16(*a0, *(a0 + 1)); + const int16x8x2_t b1 = vtrnq_s16(*(a0 + 2), *(a0 + 3)); + const int16x8x2_t b2 = vtrnq_s16(*(a0 + 4), *(a0 + 5)); + const int16x8x2_t b3 = vtrnq_s16(*(a0 + 6), *(a0 + 7)); + + // Swap 32 bit elements resulting in: + // c0.val[0]: 00 10 20 30 04 14 24 34 + // c0.val[1]: 02 12 22 32 06 16 26 36 + // c1.val[0]: 01 11 21 31 05 15 25 35 + // c1.val[1]: 03 13 23 33 07 17 27 37 + // c2.val[0]: 40 50 60 70 44 54 64 74 + // c2.val[1]: 42 52 62 72 46 56 66 76 + // c3.val[0]: 41 51 61 71 45 55 65 75 + // c3.val[1]: 43 53 63 73 47 57 67 77 + + const int32x4x2_t c0 = vtrnq_s32(vreinterpretq_s32_s16(b0.val[0]), + vreinterpretq_s32_s16(b1.val[0])); + const int32x4x2_t c1 = vtrnq_s32(vreinterpretq_s32_s16(b0.val[1]), + vreinterpretq_s32_s16(b1.val[1])); + const int32x4x2_t c2 = vtrnq_s32(vreinterpretq_s32_s16(b2.val[0]), + vreinterpretq_s32_s16(b3.val[0])); + const int32x4x2_t c3 = vtrnq_s32(vreinterpretq_s32_s16(b2.val[1]), + vreinterpretq_s32_s16(b3.val[1])); + + // Swap 64 bit elements resulting in: + // d0.val[0]: 00 10 20 30 40 50 60 70 + // d0.val[1]: 04 14 24 34 44 54 64 74 + // d1.val[0]: 01 11 21 31 41 51 61 71 + // d1.val[1]: 05 15 25 35 45 55 65 75 + // d2.val[0]: 02 12 22 32 42 52 62 72 + // d2.val[1]: 06 16 26 36 46 56 66 76 + // d3.val[0]: 03 13 23 33 43 53 63 73 + // d3.val[1]: 07 17 27 37 47 57 67 77 + const int16x8x2_t d0 = vpx_vtrnq_s64_to_s16(c0.val[0], c2.val[0]); + const int16x8x2_t d1 = vpx_vtrnq_s64_to_s16(c1.val[0], c3.val[0]); + const int16x8x2_t d2 = vpx_vtrnq_s64_to_s16(c0.val[1], c2.val[1]); + const int16x8x2_t d3 = vpx_vtrnq_s64_to_s16(c1.val[1], c3.val[1]); + + *out = d0.val[0]; + *(out + 1) = d1.val[0]; + *(out + 2) = d2.val[0]; + *(out + 3) = d3.val[0]; + *(out + 4) = d0.val[1]; + *(out + 5) = d1.val[1]; + *(out + 6) = d2.val[1]; + *(out + 7) = d3.val[1]; +} + static INLINE void transpose_s16_4x4d(int16x4_t *a0, int16x4_t *a1, int16x4_t *a2, int16x4_t *a3) { // Swap 16 bit elements. Goes from: @@ -457,4 +534,4 @@ static INLINE void transpose_s32_4x4(int32x4_t *a0, int32x4_t *a1, *a3 = c1.val[1]; } -#endif // AV1_COMMON_ARM_TRANSPOSE_NEON_H_ +#endif // AOM_AV1_COMMON_ARM_TRANSPOSE_NEON_H_ |