ref: c8f25fa5c02155367c2426eb8e4101050e2c385a
parent: 27e1bacdb3ab3317c0a1c8f62fec3f5b0824cf99
author: Linfeng Zhang <linfengz@google.com>
date: Wed Dec 14 05:42:01 EST 2016
Clean hbd idct 4x4 neon functions and other BUG=webm:1301 Change-Id: I387b7eae716a7df15c691dc6f368b07602df7342
--- a/test/partial_idct_test.cc
+++ b/test/partial_idct_test.cc
@@ -448,13 +448,13 @@
#if CONFIG_VP9_HIGHBITDEPTH
make_tuple(&vpx_highbd_fdct4x4_c,
&highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 1, 8, 2),
- make_tuple(&vpx_highbd_fdct4x4_c,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 1, 10, 2),
- make_tuple(&vpx_highbd_fdct4x4_c,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 1, 12, 2),
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 8, 2),
+ make_tuple(
+ &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 10, 2),
+ make_tuple(
+ &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 12, 2),
make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
&highbd_wrapper<vpx_highbd_idct4x4_1_add_neon>, TX_4X4, 1, 8, 2),
make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
@@ -543,13 +543,13 @@
&highbd_wrapper<vpx_highbd_idct8x8_12_add_sse2>, TX_8X8, 12, 12, 2),
make_tuple(&vpx_highbd_fdct4x4_c,
&highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 1, 8, 2),
- make_tuple(&vpx_highbd_fdct4x4_c,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 1, 10, 2),
- make_tuple(&vpx_highbd_fdct4x4_c,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
- &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 1, 12, 2),
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 8, 2),
+ make_tuple(
+ &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 10, 2),
+ make_tuple(
+ &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
+ &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 12, 2),
#endif // CONFIG_VP9_HIGHBITDEPTH
make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
&wrapper<vpx_idct32x32_1024_add_sse2>, TX_32X32, 1024, 8, 1),
@@ -620,7 +620,7 @@
make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
&wrapper<vpx_idct8x8_64_add_msa>, TX_8X8, 64, 8, 1),
make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
- &wrapper<vpx_idct8x8_12_add_msa>, TX_8X8, 10, 8, 1),
+ &wrapper<vpx_idct8x8_12_add_msa>, TX_8X8, 12, 8, 1),
make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
&wrapper<vpx_idct8x8_1_add_msa>, TX_8X8, 1, 8, 1),
make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
--- a/vpx_dsp/arm/highbd_idct4x4_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct4x4_add_neon.c
@@ -14,31 +14,56 @@
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/inv_txfm.h"
+static INLINE void highbd_idct4x4_1_add_kernel1(uint16_t **dest,
+ const int stride,
+ const int16x8_t res,
+ const int16x8_t max) {+ const uint16x4_t a0 = vld1_u16(*dest);
+ const uint16x4_t a1 = vld1_u16(*dest + stride);
+ const int16x8_t a = vreinterpretq_s16_u16(vcombine_u16(a0, a1));
+ // Note: In some profile tests, res is quite close to +/-32767.
+ // We use saturating addition.
+ const int16x8_t b = vqaddq_s16(res, a);
+ const int16x8_t c = vminq_s16(b, max);
+ const uint16x8_t d = vqshluq_n_s16(c, 0);
+ vst1_u16(*dest, vget_low_u16(d));
+ *dest += stride;
+ vst1_u16(*dest, vget_high_u16(d));
+ *dest += stride;
+}
+
+// res is in reverse row order
+static INLINE void highbd_idct4x4_1_add_kernel2(uint16_t **dest,
+ const int stride,
+ const int16x8_t res,
+ const int16x8_t max) {+ const uint16x4_t a0 = vld1_u16(*dest);
+ const uint16x4_t a1 = vld1_u16(*dest + stride);
+ const int16x8_t a = vreinterpretq_s16_u16(vcombine_u16(a1, a0));
+ // Note: In some profile tests, res is quite close to +/-32767.
+ // We use saturating addition.
+ const int16x8_t b = vqaddq_s16(res, a);
+ const int16x8_t c = vminq_s16(b, max);
+ const uint16x8_t d = vqshluq_n_s16(c, 0);
+ vst1_u16(*dest, vget_high_u16(d));
+ *dest += stride;
+ vst1_u16(*dest, vget_low_u16(d));
+ *dest += stride;
+}
+
void vpx_highbd_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest8,
int stride, int bd) {- int i;
const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
- const tran_low_t out0 = dct_const_round_shift(input[0] * cospi_16_64);
- const tran_low_t out1 = dct_const_round_shift(out0 * cospi_16_64);
+ const tran_low_t out0 =
+ HIGHBD_WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), bd);
+ const tran_low_t out1 =
+ HIGHBD_WRAPLOW(dct_const_round_shift(out0 * cospi_16_64), bd);
const int16_t a1 = ROUND_POWER_OF_TWO(out1, 4);
const int16x8_t dc = vdupq_n_s16(a1);
uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
- int16x8_t a;
- uint16x8_t b;
- uint16x4_t d0, d1;
- for (i = 0; i < 2; i++) {- d0 = vld1_u16(dest);
- d1 = vld1_u16(dest + stride);
- a = vreinterpretq_s16_u16(vcombine_u16(d0, d1));
- a = vaddq_s16(dc, a);
- a = vminq_s16(a, max);
- b = vqshluq_n_s16(a, 0);
- vst1_u16(dest, vget_low_u16(b));
- dest += stride;
- vst1_u16(dest, vget_high_u16(b));
- dest += stride;
- }
+ highbd_idct4x4_1_add_kernel1(&dest, stride, dc, max);
+ highbd_idct4x4_1_add_kernel1(&dest, stride, dc, max);
}
static INLINE void idct4x4_16_kernel_bd10(const int32x4_t cospis,
@@ -114,10 +139,7 @@
int32x4_t c2 = vld1q_s32(input + 8);
int32x4_t c3 = vld1q_s32(input + 12);
uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
- const uint16_t *dst = dest;
- int16x8_t a0, a1, d01, d32;
- int16x4_t d0, d1, d2, d3;
- uint16x8_t d01_u16, d32_u16;
+ int16x8_t a0, a1;
if (bd == 8) {const int16x4_t cospis = vld1_s16(kCospi);
@@ -142,36 +164,10 @@
idct4x4_16_kernel_bd12(cospis, &c0, &c1, &c2, &c3);
idct4x4_16_kernel_bd12(cospis, &c0, &c1, &c2, &c3);
}
- // Note: In some profile tests, a0 and a1 are quite close to +/-32767.
- // We use saturating narrow shift in case they could be even larger.
a0 = vcombine_s16(vqrshrn_n_s32(c0, 4), vqrshrn_n_s32(c1, 4));
a1 = vcombine_s16(vqrshrn_n_s32(c3, 4), vqrshrn_n_s32(c2, 4));
}
- d0 = vreinterpret_s16_u16(vld1_u16(dst));
- dst += stride;
- d1 = vreinterpret_s16_u16(vld1_u16(dst));
- dst += stride;
- d2 = vreinterpret_s16_u16(vld1_u16(dst));
- dst += stride;
- d3 = vreinterpret_s16_u16(vld1_u16(dst));
- d01 = vcombine_s16(d0, d1);
- d32 = vcombine_s16(d3, d2);
-
- // Note: In some profile tests, a0 and a1 is quite close to +/-32767.
- // We use saturating addition.
- d01 = vqaddq_s16(a0, d01);
- d32 = vqaddq_s16(a1, d32);
- d01 = vminq_s16(d01, max);
- d32 = vminq_s16(d32, max);
- d01_u16 = vqshluq_n_s16(d01, 0);
- d32_u16 = vqshluq_n_s16(d32, 0);
-
- vst1_u16(dest, vget_low_u16(d01_u16));
- dest += stride;
- vst1_u16(dest, vget_high_u16(d01_u16));
- dest += stride;
- vst1_u16(dest, vget_high_u16(d32_u16));
- dest += stride;
- vst1_u16(dest, vget_low_u16(d32_u16));
+ highbd_idct4x4_1_add_kernel1(&dest, stride, a0, max);
+ highbd_idct4x4_1_add_kernel2(&dest, stride, a1, max);
}
--- a/vpx_dsp/arm/idct8x8_add_neon.c
+++ b/vpx_dsp/arm/idct8x8_add_neon.c
@@ -16,10 +16,11 @@
#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/txfm_common.h"
-static INLINE void IDCT8x8_1D(const int16x4_t cospis0, const int16x4_t cospis1,
- int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
- int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
- int16x8_t *a6, int16x8_t *a7) {+static INLINE void idct8x8_1d(const int16x4_t cospis0, const int16x4_t cospis1,
+ int16x8_t *const a0, int16x8_t *const a1,
+ int16x8_t *const a2, int16x8_t *const a3,
+ int16x8_t *const a4, int16x8_t *const a5,
+ int16x8_t *const a6, int16x8_t *const a7) {const int16x4_t a0l = vget_low_s16(*a0);
const int16x4_t a0h = vget_high_s16(*a0);
const int16x4_t a1l = vget_low_s16(*a1);
@@ -220,17 +221,18 @@
a7 = load_tran_low_to_s16q(input + 56);
transpose_s16_8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
- IDCT8x8_1D(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+ idct8x8_1d(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
transpose_s16_8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
- IDCT8x8_1D(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+ idct8x8_1d(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
add8x8(a0, a1, a2, a3, a4, a5, a6, a7, dest, stride);
}
-static INLINE void IDCT8x4_1D(const int16x4_t cospis0, const int16x4_t cospisd0,
- const int16x4_t cospisd1, int16x8_t *a0,
- int16x8_t *a1, int16x8_t *a2, int16x8_t *a3,
- int16x8_t *a4, int16x8_t *a5, int16x8_t *a6,
- int16x8_t *a7) {+static INLINE void idct8x4_1d(const int16x4_t cospis0, const int16x4_t cospisd0,
+ const int16x4_t cospisd1, int16x8_t *const a0,
+ int16x8_t *const a1, int16x8_t *const a2,
+ int16x8_t *const a3, int16x8_t *const a4,
+ int16x8_t *const a5, int16x8_t *const a6,
+ int16x8_t *const a7) {int32x4_t b0, b1, b2, b3;
int16x4_t c0, c1, c2, c3;
int16x8_t d0, d1, d2, d3, d4, d5, d6, d7, e0, e1, e2, e3;
@@ -339,7 +341,7 @@
b7 = vsub_s16(b0, b7);
transpose_s16_4x8(b8, b9, b10, b11, b4, b5, b6, b7, &a0, &a1, &a2, &a3);
- IDCT8x4_1D(cospis0, cospisd0, cospisd1, &a0, &a1, &a2, &a3, &a4, &a5, &a6,
+ idct8x4_1d(cospis0, cospisd0, cospisd1, &a0, &a1, &a2, &a3, &a4, &a5, &a6,
&a7);
add8x8(a0, a1, a2, a3, a4, a5, a6, a7, dest, stride);
}
--- a/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
+++ b/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
@@ -263,7 +263,7 @@
RET
-; inverse 8x8 2D-DCT transform with only first 10 coeffs non-zero
+; inverse 8x8 2D-DCT transform with only first 12 coeffs non-zero
cglobal idct8x8_12_add, 3, 5, 13, input, output, stride
mova m8, [pd_8192]
mova m11, [pw_16]