Commit 7f7dd08a authored by Monty Montgomery's avatar Monty Montgomery Committed by Christopher Montgomery

Remove dead av1_dct8x8_quant_xxxx functions

They're unused, disabled in the prototype setup, but still built and
complicating the already convoluted ifdef mess in TX experiment
configuration.

Don't leave dead code in the sourcebase.  That's what SCM is for.

Change-Id: Idb2adf597ac064c7b5027df8af1cf65054984aa4
parent 0c0fd1e5
......@@ -189,9 +189,6 @@ set(AOM_AV1_ENCODER_INTRIN_SSE2
set(AOM_AV1_ENCODER_ASM_SSSE3_X86_64
"${AOM_ROOT}/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm")
set(AOM_AV1_ENCODER_INTRIN_SSSE3
"${AOM_ROOT}/av1/encoder/x86/dct_ssse3.c")
set(AOM_AV1_ENCODER_INTRIN_SSE4_1
${AOM_AV1_ENCODER_INTRIN_SSE4_1}
"${AOM_ROOT}/av1/encoder/x86/av1_highbd_quantize_sse4.c"
......@@ -221,7 +218,6 @@ if (CONFIG_HIGHBITDEPTH)
else ()
set(AOM_AV1_COMMON_INTRIN_NEON
${AOM_AV1_COMMON_INTRIN_NEON}
"${AOM_ROOT}/av1/encoder/arm/neon/dct_neon.c"
"${AOM_ROOT}/av1/common/arm/neon/iht4x4_add_neon.c"
"${AOM_ROOT}/av1/common/arm/neon/iht8x8_add_neon.c")
......@@ -468,9 +464,6 @@ if (CONFIG_PVQ)
${AOM_AV1_DECODER_INTRIN_SSE2}
"${AOM_ROOT}/av1/encoder/x86/dct_intrin_sse2.c")
set(AOM_AV1_DECODER_INTRIN_SSSE3
${AOM_AV1_DECODER_INTRIN_SSSE3}
"${AOM_ROOT}/av1/encoder/x86/dct_ssse3.c")
endif ()
endif ()
......
......@@ -146,7 +146,6 @@ AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/av1_quantize_ssse3_x86_64.asm
endif
AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_intrin_sse2.c
AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/dct_ssse3.c
AV1_CX_SRCS-$(HAVE_AVX2) += encoder/x86/hybrid_fwd_txfm_avx2.c
AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/av1_highbd_quantize_sse4.c
......@@ -161,7 +160,6 @@ endif
AV1_CX_SRCS-$(HAVE_AVX2) += encoder/x86/error_intrin_avx2.c
ifneq ($(CONFIG_HIGHBITDEPTH),yes)
AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/dct_neon.c
AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/error_neon.c
endif
AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/quantize_neon.c
......
......@@ -56,11 +56,6 @@ AV1_DX_SRCS-yes += encoder/hybrid_fwd_txfm.h
AV1_DX_SRCS-yes += encoder/dct.c
AV1_DX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
AV1_DX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_intrin_sse2.c
AV1_DX_SRCS-$(HAVE_SSSE3) += encoder/x86/dct_ssse3.c
ifneq ($(CONFIG_HIGHBITDEPTH),yes)
AV1_DX_SRCS-$(HAVE_NEON) += encoder/arm/neon/dct_neon.c
endif
AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct4x4_msa.c
AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct8x8_msa.c
......
......@@ -328,24 +328,6 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
}
#Unused
#if (aom_config("CONFIG_AOM_QM") eq "yes") {
# if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
# add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
# } else {
# add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
# }
#} else {
# if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
#
# add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
# } else {
# add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
# specialize qw/av1_fdct8x8_quant sse2 ssse3 neon/;
# }
#
#}
# fdct functions
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
......
/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <arm_neon.h>
#include "./av1_rtcd.h"
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "av1/common/blockd.h"
#include "aom_dsp/txfm_common.h"
void av1_fdct8x8_quant_neon(const int16_t *input, int stride,
int16_t *coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) {
int16_t temp_buffer[64];
(void)coeff_ptr;
aom_fdct8x8_neon(input, temp_buffer, stride);
av1_quantize_fp_neon(temp_buffer, n_coeffs, skip_block, zbin_ptr, round_ptr,
quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
}
......@@ -2015,119 +2015,6 @@ void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride,
// Note: overall scale factor of transform is 4 times unitary
}
void av1_fdct8x8_quant_c(const int16_t *input, int stride,
tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan,
const int16_t *iscan, const qm_val_t *qm_ptr,
const qm_val_t *iqm_ptr) {
int eob = -1;
int i, j;
tran_low_t intermediate[64];
// Transform columns
{
tran_low_t *output = intermediate;
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
tran_high_t t0, t1, t2, t3; // needs32
tran_high_t x0, x1, x2, x3; // canbe16
for (i = 0; i < 8; i++) {
// stage 1
s0 = (input[0 * stride] + input[7 * stride]) * 4;
s1 = (input[1 * stride] + input[6 * stride]) * 4;
s2 = (input[2 * stride] + input[5 * stride]) * 4;
s3 = (input[3 * stride] + input[4 * stride]) * 4;
s4 = (input[3 * stride] - input[4 * stride]) * 4;
s5 = (input[2 * stride] - input[5 * stride]) * 4;
s6 = (input[1 * stride] - input[6 * stride]) * 4;
s7 = (input[0 * stride] - input[7 * stride]) * 4;
// fdct4(step, step);
x0 = s0 + s3;
x1 = s1 + s2;
x2 = s1 - s2;
x3 = s0 - s3;
t0 = (x0 + x1) * cospi_16_64;
t1 = (x0 - x1) * cospi_16_64;
t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
output[0 * 8] = (tran_low_t)fdct_round_shift(t0);
output[2 * 8] = (tran_low_t)fdct_round_shift(t2);
output[4 * 8] = (tran_low_t)fdct_round_shift(t1);
output[6 * 8] = (tran_low_t)fdct_round_shift(t3);
// stage 2
t0 = (s6 - s5) * cospi_16_64;
t1 = (s6 + s5) * cospi_16_64;
t2 = fdct_round_shift(t0);
t3 = fdct_round_shift(t1);
// stage 3
x0 = s4 + t2;
x1 = s4 - t2;
x2 = s7 - t3;
x3 = s7 + t3;
// stage 4
t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
output[1 * 8] = (tran_low_t)fdct_round_shift(t0);
output[3 * 8] = (tran_low_t)fdct_round_shift(t2);
output[5 * 8] = (tran_low_t)fdct_round_shift(t1);
output[7 * 8] = (tran_low_t)fdct_round_shift(t3);
input++;
output++;
}
}
// Rows
for (i = 0; i < 8; ++i) {
fdct8(&intermediate[i * 8], &coeff_ptr[i * 8]);
for (j = 0; j < 8; ++j) coeff_ptr[j + i * 8] /= 2;
}
// TODO(jingning) Decide the need of these arguments after the
// quantization process is completed.
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)iscan;
memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
if (!skip_block) {
// Quantization pass: All coefficients with index >= zero_flag are
// skippable. Note: zero_flag can be zero.
for (i = 0; i < n_coeffs; i++) {
const int rc = scan[i];
const int coeff = coeff_ptr[rc];
const qm_val_t wt = qm_ptr != NULL ? qm_ptr[rc] : (1 << AOM_QM_BITS);
const qm_val_t iwt = iqm_ptr != NULL ? iqm_ptr[rc] : (1 << AOM_QM_BITS);
const int dequant =
(dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >>
AOM_QM_BITS;
const int coeff_sign = (coeff >> 31);
const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
int64_t tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
int tmp32;
tmp32 = (int)((tmp * quant_ptr[rc != 0] * wt) >> (16 + AOM_QM_BITS));
qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant;
if (tmp32) eob = i;
}
}
*eob_ptr = eob + 1;
}
void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride,
TxfmParam *txfm_param) {
int tx_type = txfm_param->tx_type;
......
......@@ -308,447 +308,6 @@ void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride,
}
}
void av1_fdct8x8_quant_sse2(const int16_t *input, int stride,
int16_t *coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) {
__m128i zero;
int pass;
// Constants
// When we use them, in one case, they are all the same. In all others
// it's a pair of them that we need to repeat four times. This is done
// by constructing the 32 bit constant corresponding to that pair.
const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
// Load input
__m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
__m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
__m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
__m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
__m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
__m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
__m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
__m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
__m128i *in[8];
int index = 0;
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)coeff_ptr;
// Pre-condition input (shift by two)
in0 = _mm_slli_epi16(in0, 2);
in1 = _mm_slli_epi16(in1, 2);
in2 = _mm_slli_epi16(in2, 2);
in3 = _mm_slli_epi16(in3, 2);
in4 = _mm_slli_epi16(in4, 2);
in5 = _mm_slli_epi16(in5, 2);
in6 = _mm_slli_epi16(in6, 2);
in7 = _mm_slli_epi16(in7, 2);
in[0] = &in0;
in[1] = &in1;
in[2] = &in2;
in[3] = &in3;
in[4] = &in4;
in[5] = &in5;
in[6] = &in6;
in[7] = &in7;
// We do two passes, first the columns, then the rows. The results of the
// first pass are transposed so that the same column code can be reused. The
// results of the second pass are also transposed so that the rows (processed
// as columns) are put back in row positions.
for (pass = 0; pass < 2; pass++) {
// To store results of each pass before the transpose.
__m128i res0, res1, res2, res3, res4, res5, res6, res7;
// Add/subtract
const __m128i q0 = _mm_add_epi16(in0, in7);
const __m128i q1 = _mm_add_epi16(in1, in6);
const __m128i q2 = _mm_add_epi16(in2, in5);
const __m128i q3 = _mm_add_epi16(in3, in4);
const __m128i q4 = _mm_sub_epi16(in3, in4);
const __m128i q5 = _mm_sub_epi16(in2, in5);
const __m128i q6 = _mm_sub_epi16(in1, in6);
const __m128i q7 = _mm_sub_epi16(in0, in7);
// Work on first four results
{
// Add/subtract
const __m128i r0 = _mm_add_epi16(q0, q3);
const __m128i r1 = _mm_add_epi16(q1, q2);
const __m128i r2 = _mm_sub_epi16(q1, q2);
const __m128i r3 = _mm_sub_epi16(q0, q3);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res0 = _mm_packs_epi32(w0, w1);
res4 = _mm_packs_epi32(w2, w3);
res2 = _mm_packs_epi32(w4, w5);
res6 = _mm_packs_epi32(w6, w7);
}
// Work on next four results
{
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
// dct_const_round_shift
const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
// Combine
const __m128i r0 = _mm_packs_epi32(s0, s1);
const __m128i r1 = _mm_packs_epi32(s2, s3);
// Add/subtract
const __m128i x0 = _mm_add_epi16(q4, r0);
const __m128i x1 = _mm_sub_epi16(q4, r0);
const __m128i x2 = _mm_sub_epi16(q7, r1);
const __m128i x3 = _mm_add_epi16(q7, r1);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res1 = _mm_packs_epi32(w0, w1);
res7 = _mm_packs_epi32(w2, w3);
res5 = _mm_packs_epi32(w4, w5);
res3 = _mm_packs_epi32(w6, w7);
}
// Transpose the 8x8.
{
// 00 01 02 03 04 05 06 07
// 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27
// 30 31 32 33 34 35 36 37
// 40 41 42 43 44 45 46 47
// 50 51 52 53 54 55 56 57
// 60 61 62 63 64 65 66 67
// 70 71 72 73 74 75 76 77
const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
// 04 14 05 15 06 16 07 17
// 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53
// 60 70 61 71 62 72 63 73
// 54 54 55 55 56 56 57 57
// 64 74 65 75 66 76 67 77
const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
// 00 10 20 30 01 11 21 31
// 40 50 60 70 41 51 61 71
// 02 12 22 32 03 13 23 33
// 42 52 62 72 43 53 63 73
// 04 14 24 34 05 15 21 36
// 44 54 64 74 45 55 61 76
// 06 16 26 36 07 17 27 37
// 46 56 66 76 47 57 67 77
in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
}
}
// Post-condition output and store it
{
// Post-condition (division by two)
// division of two 16 bits signed numbers using shifts
// n / 2 = (n - (n >> 15)) >> 1
const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
in0 = _mm_sub_epi16(in0, sign_in0);
in1 = _mm_sub_epi16(in1, sign_in1);
in2 = _mm_sub_epi16(in2, sign_in2);
in3 = _mm_sub_epi16(in3, sign_in3);
in4 = _mm_sub_epi16(in4, sign_in4);
in5 = _mm_sub_epi16(in5, sign_in5);
in6 = _mm_sub_epi16(in6, sign_in6);
in7 = _mm_sub_epi16(in7, sign_in7);
in0 = _mm_srai_epi16(in0, 1);
in1 = _mm_srai_epi16(in1, 1);
in2 = _mm_srai_epi16(in2, 1);
in3 = _mm_srai_epi16(in3, 1);
in4 = _mm_srai_epi16(in4, 1);
in5 = _mm_srai_epi16(in5, 1);
in6 = _mm_srai_epi16(in6, 1);
in7 = _mm_srai_epi16(in7, 1);
}
iscan_ptr += n_coeffs;
qcoeff_ptr += n_coeffs;
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
if (!skip_block) {
__m128i eob;
__m128i round, quant, dequant;
{
__m128i coeff0, coeff1;
// Setup global values
{
round = _mm_load_si128((const __m128i *)round_ptr);
quant = _mm_load_si128((const __m128i *)quant_ptr);
dequant = _mm_load_si128((const __m128i *)dequant_ptr);
}
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
// Do DC and first 15 AC
coeff0 = *in[0];
coeff1 = *in[1];
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
_mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
_mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
_mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
_mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
}
// AC only loop
index = 2;
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
coeff0 = *in[index];
coeff1 = *in[index + 1];
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);