diff options
author | Salome Thirot <salome.thirot@arm.com> | 2023-02-22 17:27:56 +0000 |
---|---|---|
committer | Salome Thirot <salome.thirot@arm.com> | 2023-02-24 11:09:57 +0000 |
commit | 6ec45f933c6c4de3fcd9344852bde25d30613321 (patch) | |
tree | 139e3e6cc8e9de8cec247ee04cb039980e1779a1 /vpx_dsp | |
parent | aab93ee6b62fd3ab489784062ea6f825f2b871da (diff) | |
download | libvpx-6ec45f933c6c4de3fcd9344852bde25d30613321.tar libvpx-6ec45f933c6c4de3fcd9344852bde25d30613321.tar.gz libvpx-6ec45f933c6c4de3fcd9344852bde25d30613321.tar.bz2 libvpx-6ec45f933c6c4de3fcd9344852bde25d30613321.zip |
Add Neon implementation of high bitdepth 16x16 hadamard transform
Add Neon implementation of vpx_highbd_hadamard_16x16 as well as the
corresponding tests.
Change-Id: If3299fe556351dfe3db994ac171d83a95ea1504b
Diffstat (limited to 'vpx_dsp')
-rw-r--r-- | vpx_dsp/arm/highbd_hadamard_neon.c | 39 | ||||
-rw-r--r-- | vpx_dsp/arm/mem_neon.h | 4 | ||||
-rw-r--r-- | vpx_dsp/vpx_dsp_rtcd_defs.pl | 2 |
3 files changed, 44 insertions, 1 deletions
diff --git a/vpx_dsp/arm/highbd_hadamard_neon.c b/vpx_dsp/arm/highbd_hadamard_neon.c index 615de4b0c..013f7148f 100644 --- a/vpx_dsp/arm/highbd_hadamard_neon.c +++ b/vpx_dsp/arm/highbd_hadamard_neon.c @@ -135,3 +135,42 @@ void vpx_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff + 32); } + +void vpx_highbd_hadamard_16x16_neon(const int16_t *src_diff, + ptrdiff_t src_stride, tran_low_t *coeff) { + int i = 0; + + // Rearrange 16x16 to 8x32 and remove stride. + // Top left first. + vpx_highbd_hadamard_8x8_neon(src_diff, src_stride, coeff); + // Top right. + vpx_highbd_hadamard_8x8_neon(src_diff + 8, src_stride, coeff + 64); + // Bottom left. + vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride, src_stride, + coeff + 128); + // Bottom right. + vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride + 8, src_stride, + coeff + 192); + + do { + int32x4_t a0 = load_tran_low_to_s32q(coeff + 4 * i); + int32x4_t a1 = load_tran_low_to_s32q(coeff + 4 * i + 64); + int32x4_t a2 = load_tran_low_to_s32q(coeff + 4 * i + 128); + int32x4_t a3 = load_tran_low_to_s32q(coeff + 4 * i + 192); + + int32x4_t b0 = vhaddq_s32(a0, a1); + int32x4_t b1 = vhsubq_s32(a0, a1); + int32x4_t b2 = vhaddq_s32(a2, a3); + int32x4_t b3 = vhsubq_s32(a2, a3); + + int32x4_t c0 = vaddq_s32(b0, b2); + int32x4_t c1 = vaddq_s32(b1, b3); + int32x4_t c2 = vsubq_s32(b0, b2); + int32x4_t c3 = vsubq_s32(b1, b3); + + store_s32q_to_tran_low(coeff + 4 * i, c0); + store_s32q_to_tran_low(coeff + 4 * i + 64, c1); + store_s32q_to_tran_low(coeff + 4 * i + 128, c2); + store_s32q_to_tran_low(coeff + 4 * i + 192, c3); + } while (++i < 16); +} diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h index b7a363891..2122956dc 100644 --- a/vpx_dsp/arm/mem_neon.h +++ b/vpx_dsp/arm/mem_neon.h @@ -106,6 +106,10 @@ static INLINE void store_s16q_to_tran_low(tran_low_t *buf, const int16x8_t a) { static INLINE void store_s32q_to_tran_low(tran_low_t *buf, const int32x4_t a) { vst1q_s32(buf, a); } + +static INLINE int32x4_t load_tran_low_to_s32q(const tran_low_t *buf) { + return vld1q_s32(buf); +} #endif // Propagate type information to the compiler. Without this the compiler may diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index dc3cdc414..276d55baf 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -805,7 +805,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") { specialize qw/vpx_highbd_hadamard_8x8 avx2 neon/; add_proto qw/void vpx_highbd_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff"; - specialize qw/vpx_highbd_hadamard_16x16 avx2/; + specialize qw/vpx_highbd_hadamard_16x16 avx2 neon/; add_proto qw/void vpx_highbd_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff"; specialize qw/vpx_highbd_hadamard_32x32 avx2/; |