ref: c338f3635e5b259fec57b8406f1416a863a4b04b
parent: ee5cb8d87f9c69819c476e865c1eba0f7ce8477e
author: Linfeng Zhang <linfengz@google.com>
date: Thu Jun 29 12:30:32 EDT 2017
Add vpx_highbd_idct8x8_{12, 64}_add_sse4_1
BUG=webm:1412
Change-Id: I5d038b4fa842ce2f6b9bd5c8c44c70647bda9591
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -242,6 +242,7 @@
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct32x32_add_sse2.c
DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_inv_txfm_sse4.h
DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_idct4x4_add_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_idct8x8_add_sse4.c
endif # !CONFIG_VP9_HIGHBITDEPTH
ifeq ($(HAVE_NEON_ASM),yes)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -653,8 +653,8 @@
if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {specialize qw/vpx_highbd_idct4x4_16_add neon sse2 sse4_1/;
- specialize qw/vpx_highbd_idct8x8_64_add neon sse2/;
- specialize qw/vpx_highbd_idct8x8_12_add neon sse2/;
+ specialize qw/vpx_highbd_idct8x8_64_add neon sse2 sse4_1/;
+ specialize qw/vpx_highbd_idct8x8_12_add neon sse2 sse4_1/;
specialize qw/vpx_highbd_idct16x16_256_add neon sse2/;
specialize qw/vpx_highbd_idct16x16_38_add neon sse2/;
$vpx_highbd_idct16x16_38_add_sse2=vpx_highbd_idct16x16_256_add_sse2;
--- a/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
@@ -187,8 +187,8 @@
highbd_idct4_large_sse2(io);
highbd_idct4_large_sse2(io);
}
- io[0] = wraplow_16bit(io[0], io[1], _mm_set1_epi32(8));
- io[1] = wraplow_16bit(io[2], io[3], _mm_set1_epi32(8));
+ io[0] = wraplow_16bit_shift4(io[0], io[1], _mm_set1_epi32(8));
+ io[1] = wraplow_16bit_shift4(io[2], io[3], _mm_set1_epi32(8));
}
recon_and_store_4(io, dest, stride, bd);
--- a/vpx_dsp/x86/highbd_idct4x4_add_sse4.c
+++ b/vpx_dsp/x86/highbd_idct4x4_add_sse4.c
@@ -67,8 +67,8 @@
} else {highbd_idct4(io);
highbd_idct4(io);
- io[0] = wraplow_16bit(io[0], io[1], _mm_set1_epi32(8));
- io[1] = wraplow_16bit(io[2], io[3], _mm_set1_epi32(8));
+ io[0] = wraplow_16bit_shift4(io[0], io[1], _mm_set1_epi32(8));
+ io[1] = wraplow_16bit_shift4(io[2], io[3], _mm_set1_epi32(8));
}
recon_and_store_4(io, dest, stride, bd);
--- /dev/null
+++ b/vpx_dsp/x86/highbd_idct8x8_add_sse4.c
@@ -1,0 +1,280 @@
+/*
+ * Copyright (c) 2017 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 <smmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/x86/highbd_inv_txfm_sse2.h"
+#include "vpx_dsp/x86/highbd_inv_txfm_sse4.h"
+#include "vpx_dsp/x86/inv_txfm_sse2.h"
+#include "vpx_dsp/x86/inv_txfm_ssse3.h"
+#include "vpx_dsp/x86/transpose_sse2.h"
+
+static void highbd_idct8x8_half1d(__m128i *const io) {+ const __m128i cp_4q_4q =
+ _mm_setr_epi32(cospi_4_64 << 2, 0, cospi_4_64 << 2, 0);
+ const __m128i cp_8q_8q =
+ _mm_setr_epi32(cospi_8_64 << 2, 0, cospi_8_64 << 2, 0);
+ const __m128i cp_12q_12q =
+ _mm_setr_epi32(cospi_12_64 << 2, 0, cospi_12_64 << 2, 0);
+ const __m128i cp_16q_16q =
+ _mm_setr_epi32(cospi_16_64 << 2, 0, cospi_16_64 << 2, 0);
+ const __m128i cp_20q_20q =
+ _mm_setr_epi32(cospi_20_64 << 2, 0, cospi_20_64 << 2, 0);
+ const __m128i cp_24q_24q =
+ _mm_setr_epi32(cospi_24_64 << 2, 0, cospi_24_64 << 2, 0);
+ const __m128i cp_28q_28q =
+ _mm_setr_epi32(cospi_28_64 << 2, 0, cospi_28_64 << 2, 0);
+ __m128i temp1[4], temp2[4], step1[8], step2[8];
+
+ transpose_32bit_4x4x2(io, io);
+
+ // stage 1
+ step1[0] = io[0];
+ step1[2] = io[4];
+ step1[1] = io[2];
+ step1[3] = io[6];
+ multiplication_and_add_2_ssse4_1(&io[1], &io[7], &cp_28q_28q, &cp_4q_4q,
+ &step1[4], &step1[7]);
+ multiplication_and_add_2_ssse4_1(&io[5], &io[3], &cp_12q_12q, &cp_20q_20q,
+ &step1[5], &step1[6]);
+
+ // stage 2
+ temp2[0] = _mm_add_epi32(step1[0], step1[2]);
+ extend_64bit(temp2[0], temp1);
+ step2[0] = multiplication_round_shift(temp1, cp_16q_16q);
+ temp2[0] = _mm_sub_epi32(step1[0], step1[2]);
+ extend_64bit(temp2[0], temp1);
+ step2[1] = multiplication_round_shift(temp1, cp_16q_16q);
+ multiplication_and_add_2_ssse4_1(&step1[1], &step1[3], &cp_24q_24q, &cp_8q_8q,
+ &step2[2], &step2[3]);
+ step2[4] = _mm_add_epi32(step1[4], step1[5]);
+ step2[5] = _mm_sub_epi32(step1[4], step1[5]);
+ step2[6] = _mm_sub_epi32(step1[7], step1[6]);
+ step2[7] = _mm_add_epi32(step1[7], step1[6]);
+
+ // stage 3
+ step1[0] = _mm_add_epi32(step2[0], step2[3]);
+ step1[1] = _mm_add_epi32(step2[1], step2[2]);
+ step1[2] = _mm_sub_epi32(step2[1], step2[2]);
+ step1[3] = _mm_sub_epi32(step2[0], step2[3]);
+ step1[4] = step2[4];
+ temp2[0] = _mm_sub_epi32(step2[6], step2[5]);
+ extend_64bit(temp2[0], temp1);
+ step1[5] = multiplication_round_shift(temp1, cp_16q_16q);
+ temp2[0] = _mm_add_epi32(step2[6], step2[5]);
+ extend_64bit(temp2[0], temp1);
+ step1[6] = multiplication_round_shift(temp1, cp_16q_16q);
+ step1[7] = step2[7];
+
+ // stage 4
+ io[0] = _mm_add_epi32(step1[0], step1[7]);
+ io[1] = _mm_add_epi32(step1[1], step1[6]);
+ io[2] = _mm_add_epi32(step1[2], step1[5]);
+ io[3] = _mm_add_epi32(step1[3], step1[4]);
+ io[4] = _mm_sub_epi32(step1[3], step1[4]);
+ io[5] = _mm_sub_epi32(step1[2], step1[5]);
+ io[6] = _mm_sub_epi32(step1[1], step1[6]);
+ io[7] = _mm_sub_epi32(step1[0], step1[7]);
+}
+
+static void highbd_idct8x8_12_half1d(__m128i *const io) {+ const __m128i cp_28q_28q =
+ _mm_setr_epi32(cospi_28_64 << 2, 0, cospi_28_64 << 2, 0);
+ const __m128i cp_4q_4q =
+ _mm_setr_epi32(cospi_4_64 << 2, 0, cospi_4_64 << 2, 0);
+ const __m128i cp_n20q_n20q =
+ _mm_setr_epi32(-cospi_20_64 << 2, 0, -cospi_20_64 << 2, 0);
+ const __m128i cp_12q_12q =
+ _mm_setr_epi32(cospi_12_64 << 2, 0, cospi_12_64 << 2, 0);
+ const __m128i cp_16q_16q =
+ _mm_setr_epi32(cospi_16_64 << 2, 0, cospi_16_64 << 2, 0);
+ const __m128i cp_8q_8q =
+ _mm_setr_epi32(cospi_8_64 << 2, 0, cospi_8_64 << 2, 0);
+ const __m128i cp_24q_24q =
+ _mm_setr_epi32(cospi_24_64 << 2, 0, cospi_24_64 << 2, 0);
+ __m128i temp1[4], temp2[4], step1[8], step2[8];
+
+ transpose_32bit_4x4(io, io);
+
+ // stage 1
+ step1[0] = io[0];
+ step1[1] = io[2];
+ extend_64bit(io[1], temp1);
+ step1[4] = multiplication_round_shift(temp1, cp_28q_28q);
+ step1[7] = multiplication_round_shift(temp1, cp_4q_4q);
+ extend_64bit(io[3], temp1);
+ step1[5] = multiplication_round_shift(temp1, cp_n20q_n20q);
+ step1[6] = multiplication_round_shift(temp1, cp_12q_12q);
+
+ // stage 2
+ extend_64bit(step1[0], temp1);
+ step2[0] = multiplication_round_shift(temp1, cp_16q_16q);
+ extend_64bit(step1[1], temp1);
+ step2[2] = multiplication_round_shift(temp1, cp_24q_24q);
+ step2[3] = multiplication_round_shift(temp1, cp_8q_8q);
+ step2[4] = _mm_add_epi32(step1[4], step1[5]);
+ step2[5] = _mm_sub_epi32(step1[4], step1[5]);
+ step2[6] = _mm_sub_epi32(step1[7], step1[6]);
+ step2[7] = _mm_add_epi32(step1[7], step1[6]);
+
+ // stage 3
+ step1[0] = _mm_add_epi32(step2[0], step2[3]);
+ step1[1] = _mm_add_epi32(step2[0], step2[2]);
+ step1[2] = _mm_sub_epi32(step2[0], step2[2]);
+ step1[3] = _mm_sub_epi32(step2[0], step2[3]);
+ step1[4] = step2[4];
+ temp2[0] = _mm_sub_epi32(step2[6], step2[5]);
+ extend_64bit(temp2[0], temp1);
+ step1[5] = multiplication_round_shift(temp1, cp_16q_16q);
+ temp2[0] = _mm_add_epi32(step2[6], step2[5]);
+ extend_64bit(temp2[0], temp1);
+ step1[6] = multiplication_round_shift(temp1, cp_16q_16q);
+ step1[7] = step2[7];
+
+ // stage 4
+ io[0] = _mm_add_epi32(step1[0], step1[7]);
+ io[1] = _mm_add_epi32(step1[1], step1[6]);
+ io[2] = _mm_add_epi32(step1[2], step1[5]);
+ io[3] = _mm_add_epi32(step1[3], step1[4]);
+ io[4] = _mm_sub_epi32(step1[3], step1[4]);
+ io[5] = _mm_sub_epi32(step1[2], step1[5]);
+ io[6] = _mm_sub_epi32(step1[1], step1[6]);
+ io[7] = _mm_sub_epi32(step1[0], step1[7]);
+}
+
+void vpx_highbd_idct8x8_64_add_sse4_1(const tran_low_t *input, uint16_t *dest,
+ int stride, int bd) {+ __m128i io[16];
+
+ io[0] = _mm_load_si128((const __m128i *)(input + 0 * 8 + 0));
+ io[4] = _mm_load_si128((const __m128i *)(input + 0 * 8 + 4));
+ io[1] = _mm_load_si128((const __m128i *)(input + 1 * 8 + 0));
+ io[5] = _mm_load_si128((const __m128i *)(input + 1 * 8 + 4));
+ io[2] = _mm_load_si128((const __m128i *)(input + 2 * 8 + 0));
+ io[6] = _mm_load_si128((const __m128i *)(input + 2 * 8 + 4));
+ io[3] = _mm_load_si128((const __m128i *)(input + 3 * 8 + 0));
+ io[7] = _mm_load_si128((const __m128i *)(input + 3 * 8 + 4));
+
+ if (bd == 8) {+ __m128i io_short[8];
+
+ io_short[0] = _mm_packs_epi32(io[0], io[4]);
+ io_short[1] = _mm_packs_epi32(io[1], io[5]);
+ io_short[2] = _mm_packs_epi32(io[2], io[6]);
+ io_short[3] = _mm_packs_epi32(io[3], io[7]);
+ io[8] = _mm_load_si128((const __m128i *)(input + 4 * 8 + 0));
+ io[12] = _mm_load_si128((const __m128i *)(input + 4 * 8 + 4));
+ io[9] = _mm_load_si128((const __m128i *)(input + 5 * 8 + 0));
+ io[13] = _mm_load_si128((const __m128i *)(input + 5 * 8 + 4));
+ io[10] = _mm_load_si128((const __m128i *)(input + 6 * 8 + 0));
+ io[14] = _mm_load_si128((const __m128i *)(input + 6 * 8 + 4));
+ io[11] = _mm_load_si128((const __m128i *)(input + 7 * 8 + 0));
+ io[15] = _mm_load_si128((const __m128i *)(input + 7 * 8 + 4));
+ io_short[4] = _mm_packs_epi32(io[8], io[12]);
+ io_short[5] = _mm_packs_epi32(io[9], io[13]);
+ io_short[6] = _mm_packs_epi32(io[10], io[14]);
+ io_short[7] = _mm_packs_epi32(io[11], io[15]);
+
+ idct8_sse2(io_short);
+ idct8_sse2(io_short);
+ round_shift_8x8(io_short, io);
+ } else {+ __m128i temp[4];
+
+ highbd_idct8x8_half1d(io);
+
+ io[8] = _mm_load_si128((const __m128i *)(input + 4 * 8 + 0));
+ io[12] = _mm_load_si128((const __m128i *)(input + 4 * 8 + 4));
+ io[9] = _mm_load_si128((const __m128i *)(input + 5 * 8 + 0));
+ io[13] = _mm_load_si128((const __m128i *)(input + 5 * 8 + 4));
+ io[10] = _mm_load_si128((const __m128i *)(input + 6 * 8 + 0));
+ io[14] = _mm_load_si128((const __m128i *)(input + 6 * 8 + 4));
+ io[11] = _mm_load_si128((const __m128i *)(input + 7 * 8 + 0));
+ io[15] = _mm_load_si128((const __m128i *)(input + 7 * 8 + 4));
+ highbd_idct8x8_half1d(&io[8]);
+
+ temp[0] = io[4];
+ temp[1] = io[5];
+ temp[2] = io[6];
+ temp[3] = io[7];
+ io[4] = io[8];
+ io[5] = io[9];
+ io[6] = io[10];
+ io[7] = io[11];
+ highbd_idct8x8_half1d(io);
+ io[8] = temp[0];
+ io[9] = temp[1];
+ io[10] = temp[2];
+ io[11] = temp[3];
+ highbd_idct8x8_half1d(&io[8]);
+
+ io[0] = wraplow_16bit_shift5(io[0], io[8], _mm_set1_epi32(16));
+ io[1] = wraplow_16bit_shift5(io[1], io[9], _mm_set1_epi32(16));
+ io[2] = wraplow_16bit_shift5(io[2], io[10], _mm_set1_epi32(16));
+ io[3] = wraplow_16bit_shift5(io[3], io[11], _mm_set1_epi32(16));
+ io[4] = wraplow_16bit_shift5(io[4], io[12], _mm_set1_epi32(16));
+ io[5] = wraplow_16bit_shift5(io[5], io[13], _mm_set1_epi32(16));
+ io[6] = wraplow_16bit_shift5(io[6], io[14], _mm_set1_epi32(16));
+ io[7] = wraplow_16bit_shift5(io[7], io[15], _mm_set1_epi32(16));
+ }
+
+ recon_and_store_8(io, dest, stride, bd);
+}
+
+void vpx_highbd_idct8x8_12_add_sse4_1(const tran_low_t *input, uint16_t *dest,
+ int stride, int bd) {+ const __m128i zero = _mm_setzero_si128();
+ __m128i io[16];
+
+ io[0] = _mm_load_si128((const __m128i *)(input + 0 * 8 + 0));
+ io[1] = _mm_load_si128((const __m128i *)(input + 1 * 8 + 0));
+ io[2] = _mm_load_si128((const __m128i *)(input + 2 * 8 + 0));
+ io[3] = _mm_load_si128((const __m128i *)(input + 3 * 8 + 0));
+
+ if (bd == 8) {+ __m128i io_short[8];
+
+ io_short[0] = _mm_packs_epi32(io[0], zero);
+ io_short[1] = _mm_packs_epi32(io[1], zero);
+ io_short[2] = _mm_packs_epi32(io[2], zero);
+ io_short[3] = _mm_packs_epi32(io[3], zero);
+
+ idct8x8_12_add_kernel_ssse3(io_short);
+ round_shift_8x8(io_short, io);
+ } else {+ __m128i temp[4];
+
+ highbd_idct8x8_12_half1d(io);
+
+ temp[0] = io[4];
+ temp[1] = io[5];
+ temp[2] = io[6];
+ temp[3] = io[7];
+ highbd_idct8x8_12_half1d(io);
+
+ io[8] = temp[0];
+ io[9] = temp[1];
+ io[10] = temp[2];
+ io[11] = temp[3];
+ highbd_idct8x8_12_half1d(&io[8]);
+
+ io[0] = wraplow_16bit_shift5(io[0], io[8], _mm_set1_epi32(16));
+ io[1] = wraplow_16bit_shift5(io[1], io[9], _mm_set1_epi32(16));
+ io[2] = wraplow_16bit_shift5(io[2], io[10], _mm_set1_epi32(16));
+ io[3] = wraplow_16bit_shift5(io[3], io[11], _mm_set1_epi32(16));
+ io[4] = wraplow_16bit_shift5(io[4], io[12], _mm_set1_epi32(16));
+ io[5] = wraplow_16bit_shift5(io[5], io[13], _mm_set1_epi32(16));
+ io[6] = wraplow_16bit_shift5(io[6], io[14], _mm_set1_epi32(16));
+ io[7] = wraplow_16bit_shift5(io[7], io[15], _mm_set1_epi32(16));
+ }
+
+ recon_and_store_8(io, dest, stride, bd);
+}
--- a/vpx_dsp/x86/highbd_inv_txfm_sse2.h
+++ b/vpx_dsp/x86/highbd_inv_txfm_sse2.h
@@ -23,8 +23,8 @@
out[1] = _mm_unpackhi_epi32(in, in); // 2, 2, 3, 3
}
-static INLINE __m128i wraplow_16bit(const __m128i in0, const __m128i in1,
- const __m128i rounding) {+static INLINE __m128i wraplow_16bit_shift4(const __m128i in0, const __m128i in1,
+ const __m128i rounding) {__m128i temp[2];
temp[0] = _mm_add_epi32(in0, rounding);
temp[1] = _mm_add_epi32(in1, rounding);
@@ -33,6 +33,16 @@
return _mm_packs_epi32(temp[0], temp[1]);
}
+static INLINE __m128i wraplow_16bit_shift5(const __m128i in0, const __m128i in1,
+ const __m128i rounding) {+ __m128i temp[2];
+ temp[0] = _mm_add_epi32(in0, rounding);
+ temp[1] = _mm_add_epi32(in1, rounding);
+ temp[0] = _mm_srai_epi32(temp[0], 5);
+ temp[1] = _mm_srai_epi32(temp[1], 5);
+ return _mm_packs_epi32(temp[0], temp[1]);
+}
+
static INLINE __m128i dct_const_round_shift_64bit(const __m128i in) {const __m128i t = _mm_add_epi64(
in,
@@ -101,6 +111,29 @@
recon_and_store_4_dual(in[0], dest, stride, bd);
dest += 2 * stride;
recon_and_store_4_dual(in[1], dest, stride, bd);
+}
+
+static INLINE void recon_and_store_8_kernel(const __m128i in,
+ uint16_t **const dest,
+ const int stride, const int bd) {+ __m128i d;
+
+ d = _mm_load_si128((const __m128i *)(*dest));
+ d = add_clamp(d, in, bd);
+ _mm_store_si128((__m128i *)(*dest), d);
+ *dest += stride;
+}
+
+static INLINE void recon_and_store_8(const __m128i *const in, uint16_t *dest,
+ const int stride, const int bd) {+ recon_and_store_8_kernel(in[0], &dest, stride, bd);
+ recon_and_store_8_kernel(in[1], &dest, stride, bd);
+ recon_and_store_8_kernel(in[2], &dest, stride, bd);
+ recon_and_store_8_kernel(in[3], &dest, stride, bd);
+ recon_and_store_8_kernel(in[4], &dest, stride, bd);
+ recon_and_store_8_kernel(in[5], &dest, stride, bd);
+ recon_and_store_8_kernel(in[6], &dest, stride, bd);
+ recon_and_store_8_kernel(in[7], &dest, stride, bd);
}
#endif // VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
--
⑨