Commit d8b1ddce authored by Lester Lu's avatar Lester Lu

Signature changes for the LGT experiment

The input arguments of av1_fht* and av1_iht* functions (and their
HBD versions) are slightly changed. Input arguments tx_type and
bd are carried by a struct fwd_txfm_param/inv_txfm_param. This
struct is meant to later on carry other prediction information,
such as intra top/left boundaries to the transform level, so
that the choice of transforms can be more adaptive to the
prediction mode and local video content.

Change-Id: Ia42544248a51845be64b72855b642ef1fe5910a9
parent cff9171e
......@@ -16,6 +16,7 @@
#include "./av1_rtcd.h"
#include "aom_dsp/txfm_common.h"
#include "av1/common/common.h"
#include "av1/common/idct.h"
static INLINE void TRANSPOSE4X4(int16x8_t *q8s16, int16x8_t *q9s16) {
int32x4_t q8s32, q9s32;
......@@ -134,7 +135,7 @@ static INLINE void IADST4x4_1D(int16x4_t *d3s16, int16x4_t *d4s16,
}
void av1_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
int dest_stride, const INV_TXFM_PARAM *param) {
uint8x8_t d26u8, d27u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
uint32x2_t d26u32, d27u32;
......@@ -148,9 +149,10 @@ void av1_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
TRANSPOSE4X4(&q8s16, &q9s16);
int tx_type = param->tx_type;
switch (tx_type) {
case 0: // idct_idct is not supported. Fall back to C
av1_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
av1_iht4x4_16_add_c(input, dest, dest_stride, param);
return;
break;
case 1: // iadst_idct
......
......@@ -16,6 +16,7 @@
#include "./av1_rtcd.h"
#include "aom_dsp/txfm_common.h"
#include "av1/common/common.h"
#include "av1/common/idct.h"
static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
int16x8_t *q10s16, int16x8_t *q11s16,
......@@ -458,7 +459,7 @@ static INLINE void IADST8X8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
}
void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
int dest_stride, const INV_TXFM_PARAM *param) {
int i;
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
......@@ -478,9 +479,10 @@ void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
int tx_type = param->tx_type;
switch (tx_type) {
case 0: // idct_idct is not supported. Fall back to C
av1_iht8x8_64_add_c(input, dest, dest_stride, tx_type);
av1_iht8x8_64_add_c(input, dest, dest_stride, param);
return;
break;
case 1: // iadst_idct
......
......@@ -11,12 +11,15 @@ print <<EOF
#include "av1/common/filter.h"
#include "av1/common/convolve.h"
#include "av1/common/av1_txfm.h"
#include "av1/common/idct.h"
#include "av1/common/odintrin.h"
struct macroblockd;
/* Encoder forward decls */
struct macroblock;
struct fwd_txfm_param;
struct inv_txfm_param;
struct aom_variance_vtable;
struct search_site_config;
struct mv;
......@@ -63,81 +66,81 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
#
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
{
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht4x4_16_add sse2/;
add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht4x8_32_add sse2/;
add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x4_32_add sse2/;
add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x16_128_add sse2/;
add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht16x8_128_add sse2/;
add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht16x32_512_add sse2/;
add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht32x16_512_add sse2/;
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x8_64_add sse2/;
add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
specialize qw/av1_iht16x16_256_add sse2 avx2/;
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
}
} else {
{
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht4x4_16_add sse2 neon dspr2/;
add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht4x8_32_add sse2/;
add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x4_32_add sse2/;
add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x16_128_add sse2/;
add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht16x8_128_add sse2/;
add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht16x32_512_add sse2/;
add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht32x16_512_add sse2/;
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
specialize qw/av1_iht8x8_64_add sse2 neon dspr2/;
add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
specialize qw/av1_iht16x16_256_add sse2 avx2 dspr2/;
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
if (aom_config("CONFIG_EXT_TX") ne "yes") {
specialize qw/av1_iht4x4_16_add msa/;
......@@ -147,10 +150,10 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
}
}
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
if (aom_config("CONFIG_TX64X64") eq "yes") {
add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
}
if (aom_config("CONFIG_NEW_QUANT") eq "yes") {
......@@ -226,31 +229,31 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
#
# dct
#
add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd";
add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
}
#inv txfm
......@@ -353,49 +356,49 @@ if (aom_config("CONFIG_AOM_QM") eq "yes") {
# fdct functions
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht4x4 sse2/;
add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht8x8 sse2/;
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht16x16 sse2 avx2/;
add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht32x32 sse2 avx2/;
if (aom_config("CONFIG_TX64X64") eq "yes") {
add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
}
add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht4x8 sse2/;
add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht8x4 sse2/;
add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht8x16 sse2/;
add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht16x8 sse2/;
add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht16x32 sse2/;
add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht32x16 sse2/;
add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
if (aom_config("CONFIG_EXT_TX") ne "yes") {
......@@ -525,25 +528,25 @@ if (aom_config("CONFIG_PVQ") eq "yes") {
# fdct functions
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht4x4 sse2/;
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht8x8 sse2/;
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht16x16 sse2/;
add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/av1_fwht4x4 sse2/;
} else {
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht4x4 sse2 msa/;
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht8x8 sse2 msa/;
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
specialize qw/av1_fht16x16 sse2 msa/;
add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
......
This diff is collapsed.
......@@ -26,7 +26,16 @@
extern "C" {
#endif
typedef struct INV_TXFM_PARAM {
// TODO(kslu) Combine FWD_TXFM_PARAM and INV_TXFM_PARAM into a common struct.
// and move the common stuff in idct.h to av1_txfm.h or txfm_common.h
typedef struct fwd_txfm_param {
TX_TYPE tx_type;
TX_SIZE tx_size;
int lossless;
int bd;
} FWD_TXFM_PARAM;
typedef struct inv_txfm_param {
#if CONFIG_ADAPT_SCAN
const int16_t *eob_threshold;
#endif
......@@ -71,12 +80,11 @@ void av1_inverse_transform_block_facade(MACROBLOCKD *xd, int plane, int block,
void av1_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
int eob, int bd);
void av1_highbd_inv_txfm_add_4x4(const tran_low_t *input, uint8_t *dest,
int stride, int eob, int bd, TX_TYPE tx_type,
int lossless);
int stride, const INV_TXFM_PARAM *param);
void av1_highbd_inv_txfm_add_4x8(const tran_low_t *input, uint8_t *dest,
int stride, int eob, int bd, TX_TYPE tx_type);
int stride, const INV_TXFM_PARAM *param);
void av1_highbd_inv_txfm_add_8x4(const tran_low_t *input, uint8_t *dest,
int stride, int eob, int bd, TX_TYPE tx_type);
int stride, const INV_TXFM_PARAM *param);
void av1_highbd_inv_txfm_add(const tran_low_t *input, uint8_t *dest, int stride,
INV_TXFM_PARAM *inv_txfm_param);
......
......@@ -23,12 +23,13 @@
#if HAVE_DSPR2
void av1_iht16x16_256_add_dspr2(const int16_t *input, uint8_t *dest, int pitch,
int tx_type) {
FWD_TXFM_PARAM *param) {
int i, j;
DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
int16_t *outptr = out;
int16_t temp_out[16];
uint32_t pos = 45;
int tx_type = param->tx_type;
/* bit positon for extract from acc */
__asm__ __volatile__("wrdsp %[pos], 1 \n\t" : : [pos] "r"(pos));
......
......@@ -23,12 +23,13 @@
#if HAVE_DSPR2
void av1_iht4x4_16_add_dspr2(const int16_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
int dest_stride, FWD_TXFM_PARAM *param) {
int i, j;
DECLARE_ALIGNED(32, int16_t, out[4 * 4]);
int16_t *outptr = out;
int16_t temp_in[4 * 4], temp_out[4];
uint32_t pos = 45;
int tx_type = param->tx_type;
/* bit positon for extract from acc */
__asm__ __volatile__("wrdsp %[pos], 1 \n\t"
......
......@@ -16,18 +16,20 @@
#include "./av1_rtcd.h"
#include "av1/common/common.h"
#include "av1/common/blockd.h"
#include "av1/common/idct.h"
#include "aom_dsp/mips/inv_txfm_dspr2.h"
#include "aom_dsp/txfm_common.h"
#include "aom_ports/mem.h"
#if HAVE_DSPR2
void av1_iht8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
int dest_stride, FWD_TXFM_PARAM *param) {
int i, j;
DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
int16_t *outptr = out;
int16_t temp_in[8 * 8], temp_out[8];
uint32_t pos = 45;
int tx_type = param->tx_type;
/* bit positon for extract from acc */
__asm__ __volatile__("wrdsp %[pos], 1 \n\t" : : [pos] "r"(pos));
......
......@@ -12,13 +12,15 @@
#include <assert.h>
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#include "aom_dsp/mips/inv_txfm_msa.h"
void av1_iht16x16_256_add_msa(const int16_t *input, uint8_t *dst,
int32_t dst_stride, int32_t tx_type) {
int32_t dst_stride, FWD_TXFM_PARAM *param) {
int32_t i;
DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
int16_t *out_ptr = &out[0];
int32_t tx_type = param->tx_type;
switch (tx_type) {
case DCT_DCT:
......
......@@ -12,11 +12,13 @@
#include <assert.h>
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#include "aom_dsp/mips/inv_txfm_msa.h"
void av1_iht4x4_16_add_msa(const int16_t *input, uint8_t *dst,
int32_t dst_stride, int32_t tx_type) {
int32_t dst_stride, FWD_TXFM_PARAM *param) {
v8i16 in0, in1, in2, in3;
int32_t tx_type = param->tx_type;
/* load vector elements of 4x4 block */
LD4x4_SH(input, in0, in1, in2, in3);
......
......@@ -12,11 +12,13 @@
#include <assert.h>
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#include "aom_dsp/mips/inv_txfm_msa.h"
void av1_iht8x8_64_add_msa(const int16_t *input, uint8_t *dst,
int32_t dst_stride, int32_t tx_type) {
int32_t dst_stride, FWD_TXFM_PARAM *param) {
v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
int32_t tx_type = param->tx_type;
/* load vector elements of 8x8 block */
LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
......
......@@ -14,6 +14,7 @@
#include "./aom_config.h"
#include "./av1_rtcd.h"
#include "av1/common/idct.h"
#include "aom_dsp/x86/inv_txfm_common_avx2.h"
void av1_idct16_avx2(__m256i *in) {
......@@ -364,8 +365,10 @@ static void iidtx16(__m256i *in) {
#endif
void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
int stride,
const INV_TXFM_PARAM *inv_txfm_param) {
__m256i in[16];
int tx_type = inv_txfm_param->tx_type;
load_buffer_16x16(input, in);
switch (tx_type) {
......
......@@ -15,6 +15,7 @@
#include "aom_dsp/x86/txfm_common_sse2.h"
#include "aom_ports/mem.h"
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#if CONFIG_EXT_TX
static INLINE void fliplr_4x4(__m128i *in /*in[2]*/) {
......@@ -59,10 +60,11 @@ static INLINE void fliplr_16x8(__m128i *in /*in[16]*/) {
#endif
void av1_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[2];
const __m128i zero = _mm_setzero_si128();
const __m128i eight = _mm_set1_epi16(8);
int tx_type = inv_txfm_param->tx_type;
in[0] = load_input_data(input);
in[1] = load_input_data(input + 8);
......@@ -150,10 +152,11 @@ void av1_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
}
void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[8];
const __m128i zero = _mm_setzero_si128();
const __m128i final_rounding = _mm_set1_epi16(1 << 4);
int tx_type = inv_txfm_param->tx_type;
// load input data
in[0] = load_input_data(input);
......@@ -251,10 +254,12 @@ static void iidtx16_sse2(__m128i *in0, __m128i *in1) {
#endif // CONFIG_EXT_TX
void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
int stride,
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[32];
__m128i *in0 = &in[0];
__m128i *in1 = &in[16];
int tx_type = inv_txfm_param->tx_type;
load_buffer_8x16(input, in0);
input += 8;
......@@ -388,8 +393,10 @@ static INLINE void flip_buffer_lr_8x8(__m128i *in) {
#endif // CONFIG_EXT_TX
void av1_iht8x16_128_add_sse2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
int stride,
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[16];
int tx_type = inv_txfm_param->tx_type;
in[0] = load_input_data(input + 0 * 8);
in[1] = load_input_data(input + 1 * 8);
......@@ -553,8 +560,10 @@ static INLINE void write_buffer_8x8_round6(uint8_t *dest, __m128i *in,
}
void av1_iht16x8_128_add_sse2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
int stride,
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[16];
int tx_type = inv_txfm_param->tx_type;
// Transpose 16x8 input into in[]
in[0] = load_input_data(input + 0 * 16);
......@@ -713,8 +722,9 @@ static INLINE void write_buffer_8x4_round5(uint8_t *dest, __m128i *in,
}
void av1_iht8x4_32_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[8];
int tx_type = inv_txfm_param->tx_type;
in[0] = load_input_data(input + 0 * 8);
in[1] = load_input_data(input + 1 * 8);
......@@ -897,8 +907,9 @@ static INLINE void write_buffer_4x8_round5(uint8_t *dest, __m128i *in,
}
void av1_iht4x8_32_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i in[8];
int tx_type = inv_txfm_param->tx_type;
// Load rows, packed two per element of 'in'.
// We pack into the bottom half of 'in' so that the
......@@ -1119,8 +1130,10 @@ static INLINE void write_buffer_16x32_round6(uint8_t *dest, __m128i *intl,
}
void av1_iht16x32_512_add_sse2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
int stride,
const INV_TXFM_PARAM *inv_txfm_param) {
__m128i intl[16], intr[16], inbl[16], inbr[16];
int tx_type = inv_txfm_param->tx_type;