summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--test/dct_test.cc5
-rw-r--r--vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c445
-rw-r--r--vp9/common/arm/neon/vp9_iht16x16_add_neon.c21
-rw-r--r--vp9/common/arm/neon/vp9_iht_neon.h12
-rw-r--r--vp9/common/vp9_rtcd_defs.pl2
-rw-r--r--vp9/vp9_common.mk1
-rw-r--r--vpx_dsp/arm/highbd_idct16x16_add_neon.c60
-rw-r--r--vpx_dsp/arm/highbd_idct32x32_135_add_neon.c1
-rw-r--r--vpx_dsp/arm/highbd_idct32x32_34_add_neon.c1
-rw-r--r--vpx_dsp/arm/highbd_idct_neon.h112
-rw-r--r--vpx_dsp/arm/idct_neon.h56
11 files changed, 587 insertions, 129 deletions
diff --git a/test/dct_test.cc b/test/dct_test.cc
index 66b2add03..d3acf7687 100644
--- a/test/dct_test.cc
+++ b/test/dct_test.cc
@@ -401,8 +401,7 @@ class TransTestBase : public ::testing::TestWithParam<DctParam> {
EXPECT_GE(static_cast<uint32_t>(limit), error)
<< "Error: " << size_ << "x" << size_
<< " inverse transform has error " << error << " at " << w << ","
- << h << " org:" << (int)src_[h * stride_ + w]
- << " opt:" << (int)dst_[h * stride_ + w];
+ << h;
if (::testing::Test::HasFailure()) {
printf("Size: %d Transform type: %d\n", size_, tx_type_);
return;
@@ -633,6 +632,8 @@ static const FuncInfo ht_neon_func_info[] = {
2 },
{ &vp9_highbd_fht8x8_c, &highbd_iht_wrapper<vp9_highbd_iht8x8_64_add_neon>, 8,
2 },
+ { &vp9_highbd_fht16x16_c,
+ &highbd_iht_wrapper<vp9_highbd_iht16x16_256_add_neon>, 16, 2 },
#endif
{ &vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1 },
{ &vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1 },
diff --git a/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c b/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c
new file mode 100644
index 000000000..057d2e9c0
--- /dev/null
+++ b/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c
@@ -0,0 +1,445 @@
+/*
+ * Copyright (c) 2018 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vp9/common/vp9_enums.h"
+#include "vp9/common/arm/neon/vp9_iht_neon.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+#include "vpx_dsp/inv_txfm.h"
+
+// Use macros to make sure argument lane is passed in as an constant integer.
+
+#define vmull_lane_s32_dual(in, c, lane, out) \
+ do { \
+ out[0].val[0] = vmull_lane_s32(vget_low_s32(in.val[0]), c, lane); \
+ out[0].val[1] = vmull_lane_s32(vget_low_s32(in.val[1]), c, lane); \
+ out[1].val[0] = vmull_lane_s32(vget_high_s32(in.val[0]), c, lane); \
+ out[1].val[1] = vmull_lane_s32(vget_high_s32(in.val[1]), c, lane); \
+ } while (0)
+
+#define vmlal_lane_s32_dual(in, c, lane, out) \
+ do { \
+ out[0].val[0] = \
+ vmlal_lane_s32(out[0].val[0], vget_low_s32(in.val[0]), c, lane); \
+ out[0].val[1] = \
+ vmlal_lane_s32(out[0].val[1], vget_low_s32(in.val[1]), c, lane); \
+ out[1].val[0] = \
+ vmlal_lane_s32(out[1].val[0], vget_high_s32(in.val[0]), c, lane); \
+ out[1].val[1] = \
+ vmlal_lane_s32(out[1].val[1], vget_high_s32(in.val[1]), c, lane); \
+ } while (0)
+
+#define vmlsl_lane_s32_dual(in, c, lane, out) \
+ do { \
+ out[0].val[0] = \
+ vmlsl_lane_s32(out[0].val[0], vget_low_s32(in.val[0]), c, lane); \
+ out[0].val[1] = \
+ vmlsl_lane_s32(out[0].val[1], vget_low_s32(in.val[1]), c, lane); \
+ out[1].val[0] = \
+ vmlsl_lane_s32(out[1].val[0], vget_high_s32(in.val[0]), c, lane); \
+ out[1].val[1] = \
+ vmlsl_lane_s32(out[1].val[1], vget_high_s32(in.val[1]), c, lane); \
+ } while (0)
+
+static INLINE int32x4x2_t
+highbd_dct_const_round_shift_low_8(const int64x2x2_t *const in) {
+ int32x4x2_t out;
+ out.val[0] = vcombine_s32(vrshrn_n_s64(in[0].val[0], DCT_CONST_BITS),
+ vrshrn_n_s64(in[1].val[0], DCT_CONST_BITS));
+ out.val[1] = vcombine_s32(vrshrn_n_s64(in[0].val[1], DCT_CONST_BITS),
+ vrshrn_n_s64(in[1].val[1], DCT_CONST_BITS));
+ return out;
+}
+
+#define highbd_iadst_half_butterfly(in, c, lane, out) \
+ do { \
+ int64x2x2_t t[2]; \
+ vmull_lane_s32_dual(in, c, lane, t); \
+ out = highbd_dct_const_round_shift_low_8(t); \
+ } while (0)
+
+#define highbd_iadst_butterfly(in0, in1, c, lane0, lane1, s0, s1) \
+ do { \
+ vmull_lane_s32_dual(in0, c, lane0, s0); \
+ vmull_lane_s32_dual(in0, c, lane1, s1); \
+ vmlal_lane_s32_dual(in1, c, lane1, s0); \
+ vmlsl_lane_s32_dual(in1, c, lane0, s1); \
+ } while (0)
+
+static INLINE int32x4x2_t vaddq_s32_dual(const int32x4x2_t in0,
+ const int32x4x2_t in1) {
+ int32x4x2_t out;
+ out.val[0] = vaddq_s32(in0.val[0], in1.val[0]);
+ out.val[1] = vaddq_s32(in0.val[1], in1.val[1]);
+ return out;
+}
+
+static INLINE int64x2x2_t vaddq_s64_dual(const int64x2x2_t in0,
+ const int64x2x2_t in1) {
+ int64x2x2_t out;
+ out.val[0] = vaddq_s64(in0.val[0], in1.val[0]);
+ out.val[1] = vaddq_s64(in0.val[1], in1.val[1]);
+ return out;
+}
+
+static INLINE int32x4x2_t vsubq_s32_dual(const int32x4x2_t in0,
+ const int32x4x2_t in1) {
+ int32x4x2_t out;
+ out.val[0] = vsubq_s32(in0.val[0], in1.val[0]);
+ out.val[1] = vsubq_s32(in0.val[1], in1.val[1]);
+ return out;
+}
+
+static INLINE int64x2x2_t vsubq_s64_dual(const int64x2x2_t in0,
+ const int64x2x2_t in1) {
+ int64x2x2_t out;
+ out.val[0] = vsubq_s64(in0.val[0], in1.val[0]);
+ out.val[1] = vsubq_s64(in0.val[1], in1.val[1]);
+ return out;
+}
+
+static INLINE int32x4x2_t vcombine_s32_dual(const int32x2x2_t in0,
+ const int32x2x2_t in1) {
+ int32x4x2_t out;
+ out.val[0] = vcombine_s32(in0.val[0], in1.val[0]);
+ out.val[1] = vcombine_s32(in0.val[1], in1.val[1]);
+ return out;
+}
+
+static INLINE int32x4x2_t highbd_add_dct_const_round_shift_low_8(
+ const int64x2x2_t *const in0, const int64x2x2_t *const in1) {
+ const int64x2x2_t sum_lo = vaddq_s64_dual(in0[0], in1[0]);
+ const int64x2x2_t sum_hi = vaddq_s64_dual(in0[1], in1[1]);
+ int32x2x2_t out_lo, out_hi;
+
+ out_lo.val[0] = vrshrn_n_s64(sum_lo.val[0], DCT_CONST_BITS);
+ out_lo.val[1] = vrshrn_n_s64(sum_lo.val[1], DCT_CONST_BITS);
+ out_hi.val[0] = vrshrn_n_s64(sum_hi.val[0], DCT_CONST_BITS);
+ out_hi.val[1] = vrshrn_n_s64(sum_hi.val[1], DCT_CONST_BITS);
+ return vcombine_s32_dual(out_lo, out_hi);
+}
+
+static INLINE int32x4x2_t highbd_sub_dct_const_round_shift_low_8(
+ const int64x2x2_t *const in0, const int64x2x2_t *const in1) {
+ const int64x2x2_t sub_lo = vsubq_s64_dual(in0[0], in1[0]);
+ const int64x2x2_t sub_hi = vsubq_s64_dual(in0[1], in1[1]);
+ int32x2x2_t out_lo, out_hi;
+
+ out_lo.val[0] = vrshrn_n_s64(sub_lo.val[0], DCT_CONST_BITS);
+ out_lo.val[1] = vrshrn_n_s64(sub_lo.val[1], DCT_CONST_BITS);
+ out_hi.val[0] = vrshrn_n_s64(sub_hi.val[0], DCT_CONST_BITS);
+ out_hi.val[1] = vrshrn_n_s64(sub_hi.val[1], DCT_CONST_BITS);
+ return vcombine_s32_dual(out_lo, out_hi);
+}
+
+static INLINE int32x4x2_t vnegq_s32_dual(const int32x4x2_t in) {
+ int32x4x2_t out;
+ out.val[0] = vnegq_s32(in.val[0]);
+ out.val[1] = vnegq_s32(in.val[1]);
+ return out;
+}
+
+void vpx_highbd_iadst16_neon(const int32_t *input, int32_t *output,
+ uint16_t *dest, const int stride, const int bd) {
+ const int32x4_t c_1_31_5_27 =
+ create_s32x4_neon(cospi_1_64, cospi_31_64, cospi_5_64, cospi_27_64);
+ const int32x4_t c_9_23_13_19 =
+ create_s32x4_neon(cospi_9_64, cospi_23_64, cospi_13_64, cospi_19_64);
+ const int32x4_t c_17_15_21_11 =
+ create_s32x4_neon(cospi_17_64, cospi_15_64, cospi_21_64, cospi_11_64);
+ const int32x4_t c_25_7_29_3 =
+ create_s32x4_neon(cospi_25_64, cospi_7_64, cospi_29_64, cospi_3_64);
+ const int32x4_t c_4_28_20_12 =
+ create_s32x4_neon(cospi_4_64, cospi_28_64, cospi_20_64, cospi_12_64);
+ const int32x4_t c_16_n16_8_24 =
+ create_s32x4_neon(cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64);
+ int32x4x2_t in[16], out[16];
+ int32x4x2_t x[16], t[12];
+ int64x2x2_t s0[2], s1[2], s2[2], s3[2], s4[2], s5[2], s6[2], s7[2];
+ int64x2x2_t s8[2], s9[2], s10[2], s11[2], s12[2], s13[2], s14[2], s15[2];
+
+ // Load input (16x8)
+ in[0].val[0] = vld1q_s32(input);
+ in[0].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[8].val[0] = vld1q_s32(input);
+ in[8].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[1].val[0] = vld1q_s32(input);
+ in[1].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[9].val[0] = vld1q_s32(input);
+ in[9].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[2].val[0] = vld1q_s32(input);
+ in[2].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[10].val[0] = vld1q_s32(input);
+ in[10].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[3].val[0] = vld1q_s32(input);
+ in[3].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[11].val[0] = vld1q_s32(input);
+ in[11].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[4].val[0] = vld1q_s32(input);
+ in[4].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[12].val[0] = vld1q_s32(input);
+ in[12].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[5].val[0] = vld1q_s32(input);
+ in[5].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[13].val[0] = vld1q_s32(input);
+ in[13].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[6].val[0] = vld1q_s32(input);
+ in[6].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[14].val[0] = vld1q_s32(input);
+ in[14].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[7].val[0] = vld1q_s32(input);
+ in[7].val[1] = vld1q_s32(input + 4);
+ input += 8;
+ in[15].val[0] = vld1q_s32(input);
+ in[15].val[1] = vld1q_s32(input + 4);
+
+ // Transpose
+ transpose_s32_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
+ &in[7]);
+ transpose_s32_8x8(&in[8], &in[9], &in[10], &in[11], &in[12], &in[13], &in[14],
+ &in[15]);
+
+ x[0] = in[15];
+ x[1] = in[0];
+ x[2] = in[13];
+ x[3] = in[2];
+ x[4] = in[11];
+ x[5] = in[4];
+ x[6] = in[9];
+ x[7] = in[6];
+ x[8] = in[7];
+ x[9] = in[8];
+ x[10] = in[5];
+ x[11] = in[10];
+ x[12] = in[3];
+ x[13] = in[12];
+ x[14] = in[1];
+ x[15] = in[14];
+
+ // stage 1
+ highbd_iadst_butterfly(x[0], x[1], vget_low_s32(c_1_31_5_27), 0, 1, s0, s1);
+ highbd_iadst_butterfly(x[2], x[3], vget_high_s32(c_1_31_5_27), 0, 1, s2, s3);
+ highbd_iadst_butterfly(x[4], x[5], vget_low_s32(c_9_23_13_19), 0, 1, s4, s5);
+ highbd_iadst_butterfly(x[6], x[7], vget_high_s32(c_9_23_13_19), 0, 1, s6, s7);
+ highbd_iadst_butterfly(x[8], x[9], vget_low_s32(c_17_15_21_11), 0, 1, s8, s9);
+ highbd_iadst_butterfly(x[10], x[11], vget_high_s32(c_17_15_21_11), 0, 1, s10,
+ s11);
+ highbd_iadst_butterfly(x[12], x[13], vget_low_s32(c_25_7_29_3), 0, 1, s12,
+ s13);
+ highbd_iadst_butterfly(x[14], x[15], vget_high_s32(c_25_7_29_3), 0, 1, s14,
+ s15);
+
+ x[0] = highbd_add_dct_const_round_shift_low_8(s0, s8);
+ x[1] = highbd_add_dct_const_round_shift_low_8(s1, s9);
+ x[2] = highbd_add_dct_const_round_shift_low_8(s2, s10);
+ x[3] = highbd_add_dct_const_round_shift_low_8(s3, s11);
+ x[4] = highbd_add_dct_const_round_shift_low_8(s4, s12);
+ x[5] = highbd_add_dct_const_round_shift_low_8(s5, s13);
+ x[6] = highbd_add_dct_const_round_shift_low_8(s6, s14);
+ x[7] = highbd_add_dct_const_round_shift_low_8(s7, s15);
+ x[8] = highbd_sub_dct_const_round_shift_low_8(s0, s8);
+ x[9] = highbd_sub_dct_const_round_shift_low_8(s1, s9);
+ x[10] = highbd_sub_dct_const_round_shift_low_8(s2, s10);
+ x[11] = highbd_sub_dct_const_round_shift_low_8(s3, s11);
+ x[12] = highbd_sub_dct_const_round_shift_low_8(s4, s12);
+ x[13] = highbd_sub_dct_const_round_shift_low_8(s5, s13);
+ x[14] = highbd_sub_dct_const_round_shift_low_8(s6, s14);
+ x[15] = highbd_sub_dct_const_round_shift_low_8(s7, s15);
+
+ // stage 2
+ t[0] = x[0];
+ t[1] = x[1];
+ t[2] = x[2];
+ t[3] = x[3];
+ t[4] = x[4];
+ t[5] = x[5];
+ t[6] = x[6];
+ t[7] = x[7];
+ highbd_iadst_butterfly(x[8], x[9], vget_low_s32(c_4_28_20_12), 0, 1, s8, s9);
+ highbd_iadst_butterfly(x[10], x[11], vget_high_s32(c_4_28_20_12), 0, 1, s10,
+ s11);
+ highbd_iadst_butterfly(x[13], x[12], vget_low_s32(c_4_28_20_12), 1, 0, s13,
+ s12);
+ highbd_iadst_butterfly(x[15], x[14], vget_high_s32(c_4_28_20_12), 1, 0, s15,
+ s14);
+
+ x[0] = vaddq_s32_dual(t[0], t[4]);
+ x[1] = vaddq_s32_dual(t[1], t[5]);
+ x[2] = vaddq_s32_dual(t[2], t[6]);
+ x[3] = vaddq_s32_dual(t[3], t[7]);
+ x[4] = vsubq_s32_dual(t[0], t[4]);
+ x[5] = vsubq_s32_dual(t[1], t[5]);
+ x[6] = vsubq_s32_dual(t[2], t[6]);
+ x[7] = vsubq_s32_dual(t[3], t[7]);
+ x[8] = highbd_add_dct_const_round_shift_low_8(s8, s12);
+ x[9] = highbd_add_dct_const_round_shift_low_8(s9, s13);
+ x[10] = highbd_add_dct_const_round_shift_low_8(s10, s14);
+ x[11] = highbd_add_dct_const_round_shift_low_8(s11, s15);
+ x[12] = highbd_sub_dct_const_round_shift_low_8(s8, s12);
+ x[13] = highbd_sub_dct_const_round_shift_low_8(s9, s13);
+ x[14] = highbd_sub_dct_const_round_shift_low_8(s10, s14);
+ x[15] = highbd_sub_dct_const_round_shift_low_8(s11, s15);
+
+ // stage 3
+ t[0] = x[0];
+ t[1] = x[1];
+ t[2] = x[2];
+ t[3] = x[3];
+ highbd_iadst_butterfly(x[4], x[5], vget_high_s32(c_16_n16_8_24), 0, 1, s4,
+ s5);
+ highbd_iadst_butterfly(x[7], x[6], vget_high_s32(c_16_n16_8_24), 1, 0, s7,
+ s6);
+ t[8] = x[8];
+ t[9] = x[9];
+ t[10] = x[10];
+ t[11] = x[11];
+ highbd_iadst_butterfly(x[12], x[13], vget_high_s32(c_16_n16_8_24), 0, 1, s12,
+ s13);
+ highbd_iadst_butterfly(x[15], x[14], vget_high_s32(c_16_n16_8_24), 1, 0, s15,
+ s14);
+
+ x[0] = vaddq_s32_dual(t[0], t[2]);
+ x[1] = vaddq_s32_dual(t[1], t[3]);
+ x[2] = vsubq_s32_dual(t[0], t[2]);
+ x[3] = vsubq_s32_dual(t[1], t[3]);
+ x[4] = highbd_add_dct_const_round_shift_low_8(s4, s6);
+ x[5] = highbd_add_dct_const_round_shift_low_8(s5, s7);
+ x[6] = highbd_sub_dct_const_round_shift_low_8(s4, s6);
+ x[7] = highbd_sub_dct_const_round_shift_low_8(s5, s7);
+ x[8] = vaddq_s32_dual(t[8], t[10]);
+ x[9] = vaddq_s32_dual(t[9], t[11]);
+ x[10] = vsubq_s32_dual(t[8], t[10]);
+ x[11] = vsubq_s32_dual(t[9], t[11]);
+ x[12] = highbd_add_dct_const_round_shift_low_8(s12, s14);
+ x[13] = highbd_add_dct_const_round_shift_low_8(s13, s15);
+ x[14] = highbd_sub_dct_const_round_shift_low_8(s12, s14);
+ x[15] = highbd_sub_dct_const_round_shift_low_8(s13, s15);
+
+ // stage 4
+ {
+ const int32x4x2_t sum = vaddq_s32_dual(x[2], x[3]);
+ const int32x4x2_t sub = vsubq_s32_dual(x[2], x[3]);
+ highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 1, x[2]);
+ highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[3]);
+ }
+ {
+ const int32x4x2_t sum = vaddq_s32_dual(x[7], x[6]);
+ const int32x4x2_t sub = vsubq_s32_dual(x[7], x[6]);
+ highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 0, x[6]);
+ highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[7]);
+ }
+ {
+ const int32x4x2_t sum = vaddq_s32_dual(x[11], x[10]);
+ const int32x4x2_t sub = vsubq_s32_dual(x[11], x[10]);
+ highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 0, x[10]);
+ highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[11]);
+ }
+ {
+ const int32x4x2_t sum = vaddq_s32_dual(x[14], x[15]);
+ const int32x4x2_t sub = vsubq_s32_dual(x[14], x[15]);
+ highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 1, x[14]);
+ highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[15]);
+ }
+
+ out[0] = x[0];
+ out[1] = vnegq_s32_dual(x[8]);
+ out[2] = x[12];
+ out[3] = vnegq_s32_dual(x[4]);
+ out[4] = x[6];
+ out[5] = x[14];
+ out[6] = x[10];
+ out[7] = x[2];
+ out[8] = x[3];
+ out[9] = x[11];
+ out[10] = x[15];
+ out[11] = x[7];
+ out[12] = x[5];
+ out[13] = vnegq_s32_dual(x[13]);
+ out[14] = x[9];
+ out[15] = vnegq_s32_dual(x[1]);
+
+ if (output) {
+ highbd_idct16x16_store_pass1(out, output);
+ } else {
+ highbd_idct16x16_add_store(out, dest, stride, bd);
+ }
+}
+
+typedef void (*highbd_iht_1d)(const int32_t *input, int32_t *output,
+ uint16_t *dest, const int stride, const int bd);
+
+typedef struct {
+ highbd_iht_1d cols, rows; // vertical and horizontal
+} highbd_iht_2d;
+
+void vp9_highbd_iht16x16_256_add_neon(const tran_low_t *input, uint16_t *dest,
+ int stride, int tx_type, int bd) {
+ if (bd == 8) {
+ static const iht_2d IHT_16[] = {
+ { vpx_idct16x16_256_add_half1d,
+ vpx_idct16x16_256_add_half1d }, // DCT_DCT = 0
+ { vpx_iadst16x16_256_add_half1d,
+ vpx_idct16x16_256_add_half1d }, // ADST_DCT = 1
+ { vpx_idct16x16_256_add_half1d,
+ vpx_iadst16x16_256_add_half1d }, // DCT_ADST = 2
+ { vpx_iadst16x16_256_add_half1d,
+ vpx_iadst16x16_256_add_half1d } // ADST_ADST = 3
+ };
+ const iht_2d ht = IHT_16[tx_type];
+ int16_t row_output[16 * 16];
+
+ // pass 1
+ ht.rows(input, row_output, dest, stride, 1); // upper 8 rows
+ ht.rows(input + 8 * 16, row_output + 8, dest, stride, 1); // lower 8 rows
+
+ // pass 2
+ ht.cols(row_output, NULL, dest, stride, 1); // left 8 columns
+ ht.cols(row_output + 16 * 8, NULL, dest + 8, stride, 1); // right 8 columns
+ } else {
+ static const highbd_iht_2d IHT_16[] = {
+ { vpx_highbd_idct16x16_256_add_half1d,
+ vpx_highbd_idct16x16_256_add_half1d }, // DCT_DCT = 0
+ { vpx_highbd_iadst16_neon,
+ vpx_highbd_idct16x16_256_add_half1d }, // ADST_DCT = 1
+ { vpx_highbd_idct16x16_256_add_half1d,
+ vpx_highbd_iadst16_neon }, // DCT_ADST = 2
+ { vpx_highbd_iadst16_neon, vpx_highbd_iadst16_neon } // ADST_ADST = 3
+ };
+ const highbd_iht_2d ht = IHT_16[tx_type];
+ int32_t row_output[16 * 16];
+
+ // pass 1
+ ht.rows(input, row_output, dest, stride, bd); // upper 8 rows
+ ht.rows(input + 8 * 16, row_output + 8, dest, stride, bd); // lower 8 rows
+
+ // pass 2
+ ht.cols(row_output, NULL, dest, stride, bd); // left 8 columns
+ ht.cols(row_output + 8 * 16, NULL, dest + 8, stride,
+ bd); // right 8 columns
+ }
+}
diff --git a/vp9/common/arm/neon/vp9_iht16x16_add_neon.c b/vp9/common/arm/neon/vp9_iht16x16_add_neon.c
index a7d5a53c7..db72ff116 100644
--- a/vp9/common/arm/neon/vp9_iht16x16_add_neon.c
+++ b/vp9/common/arm/neon/vp9_iht16x16_add_neon.c
@@ -19,9 +19,9 @@
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/arm/transpose_neon.h"
-static void iadst16x16_256_add_half1d(const void *const input, int16_t *output,
- void *const dest, const int stride,
- const int highbd_flag) {
+void vpx_iadst16x16_256_add_half1d(const void *const input, int16_t *output,
+ void *const dest, const int stride,
+ const int highbd_flag) {
int16x8_t in[16], out[16];
const int16x4_t c_1_31_5_27 =
create_s16x4_neon(cospi_1_64, cospi_31_64, cospi_5_64, cospi_27_64);
@@ -254,24 +254,17 @@ static void iadst16x16_256_add_half1d(const void *const input, int16_t *output,
}
}
-typedef void (*iht_1d)(const void *const input, int16_t *output,
- void *const dest, const int stride,
- const int highbd_flag);
-
-typedef struct {
- iht_1d cols, rows; // vertical and horizontal
-} iht_2d;
-
void vp9_iht16x16_256_add_neon(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
static const iht_2d IHT_16[] = {
{ vpx_idct16x16_256_add_half1d,
vpx_idct16x16_256_add_half1d }, // DCT_DCT = 0
- { iadst16x16_256_add_half1d,
+ { vpx_iadst16x16_256_add_half1d,
vpx_idct16x16_256_add_half1d }, // ADST_DCT = 1
{ vpx_idct16x16_256_add_half1d,
- iadst16x16_256_add_half1d }, // DCT_ADST = 2
- { iadst16x16_256_add_half1d, iadst16x16_256_add_half1d } // ADST_ADST = 3
+ vpx_iadst16x16_256_add_half1d }, // DCT_ADST = 2
+ { vpx_iadst16x16_256_add_half1d,
+ vpx_iadst16x16_256_add_half1d } // ADST_ADST = 3
};
const iht_2d ht = IHT_16[tx_type];
int16_t row_output[16 * 16];
diff --git a/vp9/common/arm/neon/vp9_iht_neon.h b/vp9/common/arm/neon/vp9_iht_neon.h
index b09b96a4d..33e68d818 100644
--- a/vp9/common/arm/neon/vp9_iht_neon.h
+++ b/vp9/common/arm/neon/vp9_iht_neon.h
@@ -257,4 +257,16 @@ static INLINE void iadst8(int16x8_t *const io) {
io[7] = vnegq_s16(x[1]);
}
+void vpx_iadst16x16_256_add_half1d(const void *const input, int16_t *output,
+ void *const dest, const int stride,
+ const int highbd_flag);
+
+typedef void (*iht_1d)(const void *const input, int16_t *output,
+ void *const dest, const int stride,
+ const int highbd_flag);
+
+typedef struct {
+ iht_1d cols, rows; // vertical and horizontal
+} iht_2d;
+
#endif // VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index c705c5ef3..20c9ed641 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -105,7 +105,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
specialize qw/vp9_highbd_iht4x4_16_add neon sse4_1/;
specialize qw/vp9_highbd_iht8x8_64_add neon sse4_1/;
- specialize qw/vp9_highbd_iht16x16_256_add sse4_1/;
+ specialize qw/vp9_highbd_iht16x16_256_add neon sse4_1/;
}
}
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index 75a0fc1fb..d40d3c445 100644
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -85,6 +85,7 @@ VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans16_dspr2.c
else
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_highbd_iht4x4_add_neon.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_highbd_iht8x8_add_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_highbd_iht16x16_add_neon.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht4x4_add_sse4.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht8x8_add_sse4.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht16x16_add_sse4.c
diff --git a/vpx_dsp/arm/highbd_idct16x16_add_neon.c b/vpx_dsp/arm/highbd_idct16x16_add_neon.c
index 3fa2f9e28..69196e174 100644
--- a/vpx_dsp/arm/highbd_idct16x16_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct16x16_add_neon.c
@@ -11,6 +11,7 @@
#include <arm_neon.h>
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/inv_txfm.h"
@@ -515,62 +516,9 @@ static INLINE void highbd_idct16x16_add_stage7(const int32x4_t *const step2,
out[15] = vsubq_s32(step2[0], step2[15]);
}
-static INLINE void highbd_idct16x16_store_pass1(const int32x4x2_t *const out,
- int32_t *output) {
- // Save the result into output
- vst1q_s32(output + 0, out[0].val[0]);
- vst1q_s32(output + 4, out[0].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[1].val[0]);
- vst1q_s32(output + 4, out[1].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[2].val[0]);
- vst1q_s32(output + 4, out[2].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[3].val[0]);
- vst1q_s32(output + 4, out[3].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[4].val[0]);
- vst1q_s32(output + 4, out[4].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[5].val[0]);
- vst1q_s32(output + 4, out[5].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[6].val[0]);
- vst1q_s32(output + 4, out[6].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[7].val[0]);
- vst1q_s32(output + 4, out[7].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[8].val[0]);
- vst1q_s32(output + 4, out[8].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[9].val[0]);
- vst1q_s32(output + 4, out[9].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[10].val[0]);
- vst1q_s32(output + 4, out[10].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[11].val[0]);
- vst1q_s32(output + 4, out[11].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[12].val[0]);
- vst1q_s32(output + 4, out[12].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[13].val[0]);
- vst1q_s32(output + 4, out[13].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[14].val[0]);
- vst1q_s32(output + 4, out[14].val[1]);
- output += 16;
- vst1q_s32(output + 0, out[15].val[0]);
- vst1q_s32(output + 4, out[15].val[1]);
-}
-
-static void vpx_highbd_idct16x16_256_add_half1d(const int32_t *input,
- int32_t *output, uint16_t *dest,
- const int stride,
- const int bd) {
+void vpx_highbd_idct16x16_256_add_half1d(const int32_t *input, int32_t *output,
+ uint16_t *dest, const int stride,
+ const int bd) {
const int32x4_t cospi_0_8_16_24 = vld1q_s32(kCospi32 + 0);
const int32x4_t cospi_4_12_20N_28 = vld1q_s32(kCospi32 + 4);
const int32x4_t cospi_2_30_10_22 = vld1q_s32(kCospi32 + 8);
diff --git a/vpx_dsp/arm/highbd_idct32x32_135_add_neon.c b/vpx_dsp/arm/highbd_idct32x32_135_add_neon.c
index 3970a5a86..6750c1a42 100644
--- a/vpx_dsp/arm/highbd_idct32x32_135_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct32x32_135_add_neon.c
@@ -12,6 +12,7 @@
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/txfm_common.h"
diff --git a/vpx_dsp/arm/highbd_idct32x32_34_add_neon.c b/vpx_dsp/arm/highbd_idct32x32_34_add_neon.c
index 5d9063b15..f05932cec 100644
--- a/vpx_dsp/arm/highbd_idct32x32_34_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct32x32_34_add_neon.c
@@ -12,6 +12,7 @@
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/txfm_common.h"
diff --git a/vpx_dsp/arm/highbd_idct_neon.h b/vpx_dsp/arm/highbd_idct_neon.h
index 612bcf5a5..a17c5a6cc 100644
--- a/vpx_dsp/arm/highbd_idct_neon.h
+++ b/vpx_dsp/arm/highbd_idct_neon.h
@@ -359,4 +359,116 @@ static INLINE void idct8x8_64_half1d_bd12(
*io7 = vsubq_s32(step1[0], step2[7]);
}
+static INLINE void highbd_idct16x16_store_pass1(const int32x4x2_t *const out,
+ int32_t *output) {
+ // Save the result into output
+ vst1q_s32(output + 0, out[0].val[0]);
+ vst1q_s32(output + 4, out[0].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[1].val[0]);
+ vst1q_s32(output + 4, out[1].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[2].val[0]);
+ vst1q_s32(output + 4, out[2].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[3].val[0]);
+ vst1q_s32(output + 4, out[3].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[4].val[0]);
+ vst1q_s32(output + 4, out[4].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[5].val[0]);
+ vst1q_s32(output + 4, out[5].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[6].val[0]);
+ vst1q_s32(output + 4, out[6].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[7].val[0]);
+ vst1q_s32(output + 4, out[7].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[8].val[0]);
+ vst1q_s32(output + 4, out[8].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[9].val[0]);
+ vst1q_s32(output + 4, out[9].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[10].val[0]);
+ vst1q_s32(output + 4, out[10].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[11].val[0]);
+ vst1q_s32(output + 4, out[11].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[12].val[0]);
+ vst1q_s32(output + 4, out[12].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[13].val[0]);
+ vst1q_s32(output + 4, out[13].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[14].val[0]);
+ vst1q_s32(output + 4, out[14].val[1]);
+ output += 16;
+ vst1q_s32(output + 0, out[15].val[0]);
+ vst1q_s32(output + 4, out[15].val[1]);
+}
+
+static INLINE void highbd_idct16x16_add_store(const int32x4x2_t *const out,
+ uint16_t *dest, const int stride,
+ const int bd) {
+ // Add the result to dest
+ const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
+ int16x8_t o[16];
+ o[0] = vcombine_s16(vrshrn_n_s32(out[0].val[0], 6),
+ vrshrn_n_s32(out[0].val[1], 6));
+ o[1] = vcombine_s16(vrshrn_n_s32(out[1].val[0], 6),
+ vrshrn_n_s32(out[1].val[1], 6));
+ o[2] = vcombine_s16(vrshrn_n_s32(out[2].val[0], 6),
+ vrshrn_n_s32(out[2].val[1], 6));
+ o[3] = vcombine_s16(vrshrn_n_s32(out[3].val[0], 6),
+ vrshrn_n_s32(out[3].val[1], 6));
+ o[4] = vcombine_s16(vrshrn_n_s32(out[4].val[0], 6),
+ vrshrn_n_s32(out[4].val[1], 6));
+ o[5] = vcombine_s16(vrshrn_n_s32(out[5].val[0], 6),
+ vrshrn_n_s32(out[5].val[1], 6));
+ o[6] = vcombine_s16(vrshrn_n_s32(out[6].val[0], 6),
+ vrshrn_n_s32(out[6].val[1], 6));
+ o[7] = vcombine_s16(vrshrn_n_s32(out[7].val[0], 6),
+ vrshrn_n_s32(out[7].val[1], 6));
+ o[8] = vcombine_s16(vrshrn_n_s32(out[8].val[0], 6),
+ vrshrn_n_s32(out[8].val[1], 6));
+ o[9] = vcombine_s16(vrshrn_n_s32(out[9].val[0], 6),
+ vrshrn_n_s32(out[9].val[1], 6));
+ o[10] = vcombine_s16(vrshrn_n_s32(out[10].val[0], 6),
+ vrshrn_n_s32(out[10].val[1], 6));
+ o[11] = vcombine_s16(vrshrn_n_s32(out[11].val[0], 6),
+ vrshrn_n_s32(out[11].val[1], 6));
+ o[12] = vcombine_s16(vrshrn_n_s32(out[12].val[0], 6),
+ vrshrn_n_s32(out[12].val[1], 6));
+ o[13] = vcombine_s16(vrshrn_n_s32(out[13].val[0], 6),
+ vrshrn_n_s32(out[13].val[1], 6));
+ o[14] = vcombine_s16(vrshrn_n_s32(out[14].val[0], 6),
+ vrshrn_n_s32(out[14].val[1], 6));
+ o[15] = vcombine_s16(vrshrn_n_s32(out[15].val[0], 6),
+ vrshrn_n_s32(out[15].val[1], 6));
+ highbd_idct16x16_add8x1(o[0], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[1], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[2], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[3], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[4], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[5], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[6], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[7], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[8], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[9], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[10], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[11], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[12], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[13], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[14], max, &dest, stride);
+ highbd_idct16x16_add8x1(o[15], max, &dest, stride);
+}
+
+void vpx_highbd_idct16x16_256_add_half1d(const int32_t *input, int32_t *output,
+ uint16_t *dest, const int stride,
+ const int bd);
+
#endif // VPX_DSP_ARM_HIGHBD_IDCT_NEON_H_
diff --git a/vpx_dsp/arm/idct_neon.h b/vpx_dsp/arm/idct_neon.h
index 73dc2a4fb..628392689 100644
--- a/vpx_dsp/arm/idct_neon.h
+++ b/vpx_dsp/arm/idct_neon.h
@@ -890,62 +890,6 @@ static INLINE void highbd_add_and_store_bd8(const int16x8_t *const a,
highbd_idct16x16_add8x1_bd8(a[31], &out, stride);
}
-static INLINE void highbd_idct16x16_add_store(const int32x4x2_t *const out,
- uint16_t *dest, const int stride,
- const int bd) {
- // Add the result to dest
- const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
- int16x8_t o[16];
- o[0] = vcombine_s16(vrshrn_n_s32(out[0].val[0], 6),
- vrshrn_n_s32(out[0].val[1], 6));
- o[1] = vcombine_s16(vrshrn_n_s32(out[1].val[0], 6),
- vrshrn_n_s32(out[1].val[1], 6));
- o[2] = vcombine_s16(vrshrn_n_s32(out[2].val[0], 6),
- vrshrn_n_s32(out[2].val[1], 6));
- o[3] = vcombine_s16(vrshrn_n_s32(out[3].val[0], 6),
- vrshrn_n_s32(out[3].val[1], 6));
- o[4] = vcombine_s16(vrshrn_n_s32(out[4].val[0], 6),
- vrshrn_n_s32(out[4].val[1], 6));
- o[5] = vcombine_s16(vrshrn_n_s32(out[5].val[0], 6),
- vrshrn_n_s32(out[5].val[1], 6));
- o[6] = vcombine_s16(vrshrn_n_s32(out[6].val[0], 6),
- vrshrn_n_s32(out[6].val[1], 6));
- o[7] = vcombine_s16(vrshrn_n_s32(out[7].val[0], 6),
- vrshrn_n_s32(out[7].val[1], 6));
- o[8] = vcombine_s16(vrshrn_n_s32(out[8].val[0], 6),
- vrshrn_n_s32(out[8].val[1], 6));
- o[9] = vcombine_s16(vrshrn_n_s32(out[9].val[0], 6),
- vrshrn_n_s32(out[9].val[1], 6));
- o[10] = vcombine_s16(vrshrn_n_s32(out[10].val[0], 6),
- vrshrn_n_s32(out[10].val[1], 6));
- o[11] = vcombine_s16(vrshrn_n_s32(out[11].val[0], 6),
- vrshrn_n_s32(out[11].val[1], 6));
- o[12] = vcombine_s16(vrshrn_n_s32(out[12].val[0], 6),
- vrshrn_n_s32(out[12].val[1], 6));
- o[13] = vcombine_s16(vrshrn_n_s32(out[13].val[0], 6),
- vrshrn_n_s32(out[13].val[1], 6));
- o[14] = vcombine_s16(vrshrn_n_s32(out[14].val[0], 6),
- vrshrn_n_s32(out[14].val[1], 6));
- o[15] = vcombine_s16(vrshrn_n_s32(out[15].val[0], 6),
- vrshrn_n_s32(out[15].val[1], 6));
- highbd_idct16x16_add8x1(o[0], max, &dest, stride);
- highbd_idct16x16_add8x1(o[1], max, &dest, stride);
- highbd_idct16x16_add8x1(o[2], max, &dest, stride);
- highbd_idct16x16_add8x1(o[3], max, &dest, stride);
- highbd_idct16x16_add8x1(o[4], max, &dest, stride);
- highbd_idct16x16_add8x1(o[5], max, &dest, stride);
- highbd_idct16x16_add8x1(o[6], max, &dest, stride);
- highbd_idct16x16_add8x1(o[7], max, &dest, stride);
- highbd_idct16x16_add8x1(o[8], max, &dest, stride);
- highbd_idct16x16_add8x1(o[9], max, &dest, stride);
- highbd_idct16x16_add8x1(o[10], max, &dest, stride);
- highbd_idct16x16_add8x1(o[11], max, &dest, stride);
- highbd_idct16x16_add8x1(o[12], max, &dest, stride);
- highbd_idct16x16_add8x1(o[13], max, &dest, stride);
- highbd_idct16x16_add8x1(o[14], max, &dest, stride);
- highbd_idct16x16_add8x1(o[15], max, &dest, stride);
-}
-
void vpx_idct16x16_256_add_half1d(const void *const input, int16_t *output,
void *const dest, const int stride,
const int highbd_flag);
generated by cgit v1.2.3 (git 2.25.1) at 2024-10-09 23:34:38 +0000