Commit f1342a7b authored by Geza Lore's avatar Geza Lore

Add AVX vectorized vp9_diamond_search_sad

This function now has an AVX intrinsics version which is about 80%
faster compared to the C implementation. This provides a 2-4% total
speed-up for encode, depending on encoding parameters. The function
utilizes 3 properties of the cost function lookup table, constructed
in 'cal_nmvjointsadcost' and 'cal_nmvsadcosts'.
For the joint cost:
  - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]
For the component costs:
  - For all i: mvsadcost[0][i] == mvsadcost[1][i]
        (equal per component cost)
  - For all i: mvsadcost[0][i] == mvsadcost[0][-i]
        (Cost function is even)
These must hold, otherwise the AVX version of the function cannot be used.

Change-Id: I184055b864c5a2dc37b2d8c5c9012eb801e9daf6
parent 965a8dea
......@@ -312,7 +312,7 @@ $vp9_full_search_sad_sse3=vp9_full_search_sadx3;
$vp9_full_search_sad_sse4_1=vp9_full_search_sadx8;
add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
specialize qw/vp9_diamond_search_sad/;
specialize qw/vp9_diamond_search_sad avx/;
add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
specialize qw/vp9_full_range_search/;
......
......@@ -1567,7 +1567,30 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
#endif
#define log2f(x) (log (x) / (float) M_LOG2_E)
/***********************************************************************
* Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' *
***********************************************************************
* The following 2 functions ('cal_nmvjointsadcost' and *
* 'cal_nmvsadcosts') are used to calculate cost lookup tables *
* used by 'vp9_diamond_search_sad'. The C implementation of the *
* function is generic, but the AVX intrinsics optimised version *
* relies on the following properties of the computed tables: *
* For cal_nmvjointsadcost: *
* - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
* For cal_nmvsadcosts: *
* - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
* (Equal costs for both components) *
* - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
* (Cost function is even) *
* If these do not hold, then the AVX optimised version of the *
* 'vp9_diamond_search_sad' function cannot be used as it is, in which *
* case you can revert to using the C function instead. *
***********************************************************************/
static void cal_nmvjointsadcost(int *mvjointsadcost) {
/*********************************************************************
* Warning: Read the comments above before modifying this function *
*********************************************************************/
mvjointsadcost[0] = 600;
mvjointsadcost[1] = 300;
mvjointsadcost[2] = 300;
......@@ -1575,6 +1598,9 @@ static void cal_nmvjointsadcost(int *mvjointsadcost) {
}
static void cal_nmvsadcosts(int *mvsadcost[2]) {
/*********************************************************************
* Warning: Read the comments above before modifying this function *
*********************************************************************/
int i = 1;
mvsadcost[0][0] = 0;
......@@ -1736,6 +1762,10 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
cpi->first_time_stamp_ever = INT64_MAX;
/*********************************************************************
* Warning: Read the comments around 'cal_nmvjointsadcost' and *
* 'cal_nmvsadcosts' before modifying how these tables are computed. *
*********************************************************************/
cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
......
......@@ -101,11 +101,8 @@ static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
}
void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
int len, ss_count = 1;
cfg->ss_mv[0].col = 0;
cfg->ss_mv[0].row = 0;
cfg->ss_os[0] = 0;
int len;
int ss_count = 0;
for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
// Generate offsets for 4 search sites per step.
......@@ -117,16 +114,13 @@ void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
}
}
cfg->ss_count = ss_count;
cfg->searches_per_step = 4;
cfg->total_steps = ss_count / cfg->searches_per_step;
}
void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
int len, ss_count = 1;
cfg->ss_mv[0].col = 0;
cfg->ss_mv[0].row = 0;
cfg->ss_os[0] = 0;
int len;
int ss_count = 0;
for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
// Generate offsets for 8 search sites per step.
......@@ -141,8 +135,8 @@ void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
}
}
cfg->ss_count = ss_count;
cfg->searches_per_step = 8;
cfg->total_steps = ss_count / cfg->searches_per_step;
}
/*
......@@ -1612,8 +1606,8 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x,
const uint8_t *best_address;
unsigned int bestsad = INT_MAX;
int best_site = 0;
int last_site = 0;
int best_site = -1;
int last_site = -1;
int ref_row;
int ref_col;
......@@ -1626,7 +1620,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x,
// const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param;
const int tot_steps = (cfg->total_steps) - search_param;
const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
......@@ -1644,7 +1638,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x,
bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
+ mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
i = 1;
i = 0;
for (step = 0; step < tot_steps; step++) {
int all_in = 1, t;
......
......@@ -33,10 +33,10 @@ extern "C" {
typedef struct search_site_config {
// motion search sites
MV ss_mv[8 * MAX_MVSEARCH_STEPS + 1]; // Motion vector
intptr_t ss_os[8 * MAX_MVSEARCH_STEPS + 1]; // Offset
int ss_count;
MV ss_mv[8 * MAX_MVSEARCH_STEPS]; // Motion vector
intptr_t ss_os[8 * MAX_MVSEARCH_STEPS]; // Offset
int searches_per_step;
int total_steps;
} search_site_config;
void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride);
......
/*
* Copyright (c) 2015 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 <emmintrin.h>
#include <smmintrin.h>
#include "vpx_dsp/vpx_dsp_common.h"
#include "vp9/encoder/vp9_encoder.h"
#include "vpx_ports/mem.h"
#ifdef __GNUC__
# define __likely__(v) __builtin_expect(v, 1)
# define __unlikely__(v) __builtin_expect(v, 0)
#else
# define __likely__(v) (v)
# define __unlikely__(v) (v)
#endif
static INLINE MV_JOINT_TYPE get_mv_joint(const int_mv mv) {
// This is simplified from the C implementation to utilise that
// x->nmvjointsadcost[1] == x->nmvjointsadcost[2] and
// x->nmvjointsadcost[1] == x->nmvjointsadcost[3]
return mv.as_int == 0 ? 0 : 1;
}
static INLINE int mv_cost(const int_mv mv,
const int *joint_cost, int *const comp_cost[2]) {
return joint_cost[get_mv_joint(mv)] +
comp_cost[0][mv.as_mv.row] + comp_cost[1][mv.as_mv.col];
}
static int mvsad_err_cost(const MACROBLOCK *x, const int_mv mv, const MV *ref,
int error_per_bit) {
const int_mv diff = { .as_mv={ mv.as_mv.row - ref->row,
mv.as_mv.col - ref->col } };
return ROUND_POWER_OF_TWO(mv_cost(diff, x->nmvjointsadcost,
x->nmvsadcost) * error_per_bit, 8);
}
/*****************************************************************************
* This function utilises 3 properties of the cost function lookup tables, *
* constructed in using 'cal_nmvjointsadcost' and 'cal_nmvsadcosts' in *
* vp9_encoder.c. *
* For the joint cost: *
* - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
* For the component costs: *
* - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
* (Equal costs for both components) *
* - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
* (Cost function is even) *
* If these do not hold, then this function cannot be used without *
* modification, in which case you can revert to using the C implementation, *
* which does not rely on these properties. *
*****************************************************************************/
int vp9_diamond_search_sad_avx(const MACROBLOCK *x,
const search_site_config *cfg,
MV *ref_mv, MV *best_mv, int search_param,
int sad_per_bit, int *num00,
const vp9_variance_fn_ptr_t *fn_ptr,
const MV *center_mv) {
const int_mv maxmv = { .as_mv={x->mv_row_max, x->mv_col_max} };
const __m128i v_max_mv_w = _mm_set1_epi32(maxmv.as_int);
const int_mv minmv = { .as_mv={x->mv_row_min, x->mv_col_min} };
const __m128i v_min_mv_w = _mm_set1_epi32(minmv.as_int);
const __m128i v_spb_d = _mm_set1_epi32(sad_per_bit);
const __m128i v_joint_cost_0_d = _mm_set1_epi32(x->nmvjointsadcost[0]);
const __m128i v_joint_cost_1_d = _mm_set1_epi32(x->nmvjointsadcost[1]);
// search_param determines the length of the initial step and hence the number
// of iterations.
// 0 = initial step (MAX_FIRST_STEP) pel
// 1 = (MAX_FIRST_STEP/2) pel,
// 2 = (MAX_FIRST_STEP/4) pel...
const MV *ss_mv = &cfg->ss_mv[cfg->searches_per_step * search_param];
const intptr_t *ss_os = &cfg->ss_os[cfg->searches_per_step * search_param];
const int tot_steps = cfg->total_steps - search_param;
const int_mv fcenter_mv = { .as_mv={center_mv->row >> 3,
center_mv->col >> 3} };
const __m128i vfcmv = _mm_set1_epi32(fcenter_mv.as_int);
const int ref_row = clamp(ref_mv->row, minmv.as_mv.row, maxmv.as_mv.row);
const int ref_col = clamp(ref_mv->col, minmv.as_mv.col, maxmv.as_mv.col);
int_mv bmv = { .as_mv={ ref_row, ref_col} };
int_mv new_bmv = bmv;
__m128i v_bmv_w = _mm_set1_epi32(bmv.as_int);
const int what_stride = x->plane[0].src.stride;
const uint8_t * const what = x->plane[0].src.buf;
const int in_what_stride = x->e_mbd.plane[0].pre[0].stride;
const uint8_t * const in_what = x->e_mbd.plane[0].pre[0].buf +
ref_row * in_what_stride + ref_col;
// Work out the start point for the search
const uint8_t *best_address = in_what;
const uint8_t *new_best_address = best_address;
#if ARCH_X86_64
__m128i v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
#else
__m128i v_ba_d = _mm_set1_epi32((intptr_t)best_address);
#endif
unsigned int best_sad;
int i;
int j;
int step;
// Check the starting position
best_sad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
best_sad += mvsad_err_cost(x, bmv, &fcenter_mv.as_mv, sad_per_bit);
*num00 = 0;
for (i = 0, step = 0; step < tot_steps; step++) {
for (j = 0; j < cfg->searches_per_step; j += 4, i += 4) {
__m128i v_sad_d;
__m128i v_cost_d;
__m128i v_outside_d;
__m128i v_inside_d;
__m128i v_diff_mv_w;
#if ARCH_X86_64
__m128i v_blocka[2];
#else
__m128i v_blocka[1];
#endif
// Compute the candidate motion vectors
const __m128i v_ss_mv_w = _mm_load_si128((const __m128i*)&ss_mv[i]);
const __m128i v_these_mv_w = _mm_add_epi16(v_bmv_w, v_ss_mv_w);
// Clamp them to the search bounds
__m128i v_these_mv_clamp_w = v_these_mv_w;
v_these_mv_clamp_w = _mm_min_epi16(v_these_mv_clamp_w, v_max_mv_w);
v_these_mv_clamp_w = _mm_max_epi16(v_these_mv_clamp_w, v_min_mv_w);
// The ones that did not change are inside the search area
v_inside_d = _mm_cmpeq_epi32(v_these_mv_clamp_w, v_these_mv_w);
// If none of them are inside, then move on
if (__likely__(_mm_test_all_zeros(v_inside_d, v_inside_d))) {
continue;
}
// The inverse mask indicates which of the MVs are outside
v_outside_d = _mm_xor_si128(v_inside_d, _mm_set1_epi8(0xff));
// Shift right to keep the sign bit clear, we will use this later
// to set the cost to the maximum value.
v_outside_d = _mm_srli_epi32(v_outside_d, 1);
// Compute the difference MV
v_diff_mv_w = _mm_sub_epi16(v_these_mv_clamp_w, vfcmv);
// We utilise the fact that the cost function is even, and use the
// absolute difference. This allows us to use unsigned indexes later
// and reduces cache pressure somewhat as only a half of the table
// is ever referenced.
v_diff_mv_w = _mm_abs_epi16(v_diff_mv_w);
// Compute the SIMD pointer offsets.
{
#if ARCH_X86_64 // sizeof(intptr_t) == 8
// Load the offsets (could use _mm_maskload_ps here, instead of the
// extra 'and' but it's slower that way)
__m128i v_bo10_q = _mm_load_si128((const __m128i*)&ss_os[i+0]);
__m128i v_bo32_q = _mm_load_si128((const __m128i*)&ss_os[i+2]);
// Set the ones falling outside to zero
v_bo10_q = _mm_and_si128(v_bo10_q,
_mm_cvtepi32_epi64(v_inside_d));
v_bo32_q = _mm_and_si128(v_bo32_q,
_mm_unpackhi_epi32(v_inside_d, v_inside_d));
// Compute the candidate addresses
v_blocka[0] = _mm_add_epi64(v_ba_q, v_bo10_q);
v_blocka[1] = _mm_add_epi64(v_ba_q, v_bo32_q);
#else // ARCH_X86 // sizeof(intptr_t) == 4
__m128i v_bo_d = _mm_load_si128((const __m128i*)&ss_os[i]);
v_bo_d = _mm_and_si128(v_bo_d, v_inside_d);
v_blocka[0] = _mm_add_epi32(v_ba_d, v_bo_d);
#endif
}
fn_ptr->sdx4df(what, what_stride,
(const uint8_t **)&v_blocka[0], in_what_stride,
(uint32_t*)&v_sad_d);
// Look up the component cost of the residual motion vector
{
const int32_t row0 = _mm_extract_epi16(v_diff_mv_w, 0);
const int32_t col0 = _mm_extract_epi16(v_diff_mv_w, 1);
const int32_t row1 = _mm_extract_epi16(v_diff_mv_w, 2);
const int32_t col1 = _mm_extract_epi16(v_diff_mv_w, 3);
const int32_t row2 = _mm_extract_epi16(v_diff_mv_w, 4);
const int32_t col2 = _mm_extract_epi16(v_diff_mv_w, 5);
const int32_t row3 = _mm_extract_epi16(v_diff_mv_w, 6);
const int32_t col3 = _mm_extract_epi16(v_diff_mv_w, 7);
// Note: This is a use case for vpgather in AVX2
const uint32_t cost0 = x->nmvsadcost[0][row0] + x->nmvsadcost[0][col0];
const uint32_t cost1 = x->nmvsadcost[0][row1] + x->nmvsadcost[0][col1];
const uint32_t cost2 = x->nmvsadcost[0][row2] + x->nmvsadcost[0][col2];
const uint32_t cost3 = x->nmvsadcost[0][row3] + x->nmvsadcost[0][col3];
__m128i v_cost_10_d;
__m128i v_cost_32_d;
v_cost_10_d = _mm_cvtsi32_si128(cost0);
v_cost_10_d = _mm_insert_epi32(v_cost_10_d, cost1, 1);
v_cost_32_d = _mm_cvtsi32_si128(cost2);
v_cost_32_d = _mm_insert_epi32(v_cost_32_d, cost3, 1);
v_cost_d = _mm_unpacklo_epi64(v_cost_10_d, v_cost_32_d);
}
// Now add in the joint cost
{
const __m128i v_sel_d = _mm_cmpeq_epi32(v_diff_mv_w,
_mm_setzero_si128());
const __m128i v_joint_cost_d = _mm_blendv_epi8(v_joint_cost_1_d,
v_joint_cost_0_d,
v_sel_d);
v_cost_d = _mm_add_epi32(v_cost_d, v_joint_cost_d);
}
// Multiply by sad_per_bit
v_cost_d = _mm_mullo_epi32(v_cost_d, v_spb_d);
// ROUND_POWER_OF_TWO(v_cost_d, 8)
v_cost_d = _mm_add_epi32(v_cost_d, _mm_set1_epi32(0x80));
v_cost_d = _mm_srai_epi32(v_cost_d, 8);
// Add the cost to the sad
v_sad_d = _mm_add_epi32(v_sad_d, v_cost_d);
// Make the motion vectors outside the search area have max cost
// by or'ing in the comparison mask, this way the minimum search won't
// pick them.
v_sad_d = _mm_or_si128(v_sad_d, v_outside_d);
// Find the minimum value and index horizontally in v_sad_d
{
// Try speculatively on 16 bits, so we can use the minpos intrinsic
const __m128i v_sad_w = _mm_packus_epi32(v_sad_d, v_sad_d);
const __m128i v_minp_w = _mm_minpos_epu16(v_sad_w);
uint32_t local_best_sad = _mm_extract_epi16(v_minp_w, 0);
uint32_t local_best_idx = _mm_extract_epi16(v_minp_w, 1);
// If the local best value is not saturated, just use it, otherwise
// find the horizontal minimum again the hard way on 32 bits.
// This is executed rarely.
if (__unlikely__(local_best_sad == 0xffff)) {
__m128i v_loval_d, v_hival_d, v_loidx_d, v_hiidx_d, v_sel_d;
v_loval_d = v_sad_d;
v_loidx_d = _mm_set_epi32(3, 2, 1, 0);
v_hival_d = _mm_srli_si128(v_loval_d, 8);
v_hiidx_d = _mm_srli_si128(v_loidx_d, 8);
v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
v_hival_d = _mm_srli_si128(v_loval_d, 4);
v_hiidx_d = _mm_srli_si128(v_loidx_d, 4);
v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
local_best_sad = _mm_extract_epi32(v_loval_d, 0);
local_best_idx = _mm_extract_epi32(v_loidx_d, 0);
}
// Update the global minimum if the local minimum is smaller
if (__likely__(local_best_sad < best_sad)) {
new_bmv = ((const int_mv *)&v_these_mv_w)[local_best_idx];
new_best_address = ((const uint8_t **)v_blocka)[local_best_idx];
best_sad = local_best_sad;
}
}
}
bmv = new_bmv;
best_address = new_best_address;
v_bmv_w = _mm_set1_epi32(bmv.as_int);
#if ARCH_X86_64
v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
#else
v_ba_d = _mm_set1_epi32((intptr_t)best_address);
#endif
if (__unlikely__(best_address == in_what)) {
(*num00)++;
}
}
*best_mv = bmv.as_mv;
return best_sad;
}
......@@ -94,6 +94,7 @@ VP9_CX_SRCS-yes += encoder/vp9_mbgraph.h
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_avg_intrin_sse2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c
VP9_CX_SRCS-$(HAVE_AVX) += encoder/x86/vp9_diamond_search_sad_avx.c
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_block_error_intrin_sse2.c
endif
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment