Commit 29d47ac8 authored by John Koleszar's avatar John Koleszar
Browse files

Restore SSSE3 subpixel filters in new convolve framework

This commit adds the 8 tap SSSE3 subpixel filters back into the code
underneath the convolve API. The C code is still called for 4x4
blocks, as well as compound prediction modes. This restores the
encode performance to be within about 8% of the baseline.

Change-Id: Ife0d81477075ae33c05b53c65003951efdc8b09c
parent 7a07eea1
......@@ -12,6 +12,7 @@
extern "C" {
#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "vpx_mem/vpx_mem.h"
}
#include "third_party/googletest/src/include/gtest/gtest.h"
#include "test/acm_random.h"
......@@ -166,7 +167,25 @@ static void filter_average_block2d_8_c(const uint8_t *src_ptr,
}
class ConvolveTest : public PARAMS(int, int, const ConvolveFunctions*) {
public:
static void SetUpTestCase() {
// Force input_ to be unaligned, output to be 16 byte aligned.
input_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kOuterBlockSize * kOuterBlockSize + 1))
+ 1;
output_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kOuterBlockSize * kOuterBlockSize));
}
static void TearDownTestCase() {
vpx_free(input_ - 1);
input_ = NULL;
vpx_free(output_);
output_ = NULL;
}
protected:
static const int kDataAlignment = 16;
static const int kOuterBlockSize = 32;
static const int kInputStride = kOuterBlockSize;
static const int kOutputStride = kOuterBlockSize;
......@@ -174,7 +193,10 @@ class ConvolveTest : public PARAMS(int, int, const ConvolveFunctions*) {
int Width() const { return GET_PARAM(0); }
int Height() const { return GET_PARAM(1); }
int BorderLeft() const { return (kOuterBlockSize - Width()) / 2; }
int BorderLeft() const {
const int center = (kOuterBlockSize - Width()) / 2;
return (center + (kDataAlignment - 1)) & ~(kDataAlignment - 1);
}
int BorderTop() const { return (kOuterBlockSize - Height()) / 2; }
bool IsIndexInBorder(int i) {
......@@ -216,9 +238,11 @@ class ConvolveTest : public PARAMS(int, int, const ConvolveFunctions*) {
}
const ConvolveFunctions* UUT_;
uint8_t input_[kOuterBlockSize * kOuterBlockSize];
uint8_t output_[kOuterBlockSize * kOuterBlockSize];
static uint8_t* input_;
static uint8_t* output_;
};
uint8_t* ConvolveTest::input_ = NULL;
uint8_t* ConvolveTest::output_ = NULL;
TEST_P(ConvolveTest, GuardBlocks) {
CheckGuardBlocks();
......@@ -488,3 +512,16 @@ INSTANTIATE_TEST_CASE_P(C, ConvolveTest, ::testing::Values(
make_tuple(8, 8, &convolve8_c),
make_tuple(16, 16, &convolve8_c)));
}
#if HAVE_SSSE3
const ConvolveFunctions convolve8_ssse3(
vp9_convolve8_horiz_ssse3, vp9_convolve8_avg_horiz_c,
vp9_convolve8_vert_ssse3, vp9_convolve8_avg_vert_c,
vp9_convolve8_ssse3, vp9_convolve8_avg_c);
INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_ssse3),
make_tuple(8, 4, &convolve8_ssse3),
make_tuple(8, 8, &convolve8_ssse3),
make_tuple(16, 16, &convolve8_ssse3)));
#endif
......@@ -245,13 +245,13 @@ specialize vp9_sub_pixel_variance16x2 sse2
# Sub Pixel Filters
#
prototype void vp9_convolve8 "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
specialize vp9_convolve8
specialize vp9_convolve8 ssse3
prototype void vp9_convolve8_horiz "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
specialize vp9_convolve8_horiz
specialize vp9_convolve8_horiz ssse3
prototype void vp9_convolve8_vert "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
specialize vp9_convolve8_vert
specialize vp9_convolve8_vert ssse3
prototype void vp9_convolve8_avg "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
specialize vp9_convolve8_avg
......
......@@ -8,8 +8,10 @@
* be found in the AUTHORS file in the root of the source tree.
*/
#include <assert.h>
#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "vpx_ports/mem.h"
///////////////////////////////////////////////////////////////////////////
// the mmx function that does the bilinear filtering and var calculation //
......@@ -33,3 +35,130 @@ DECLARE_ALIGNED(16, const short, vp9_bilinear_filters_mmx[16][8]) = {
{ 16, 16, 16, 16, 112, 112, 112, 112 },
{ 8, 8, 8, 8, 120, 120, 120, 120 }
};
#if HAVE_SSSE3
void vp9_filter_block1d16_v8_ssse3(const unsigned char *src_ptr,
const unsigned int src_pitch,
unsigned char *output_ptr,
unsigned int out_pitch,
unsigned int output_height,
const short *filter);
void vp9_filter_block1d16_h8_ssse3(const unsigned char *src_ptr,
const unsigned int src_pitch,
unsigned char *output_ptr,
unsigned int out_pitch,
unsigned int output_height,
const short *filter);
void vp9_filter_block1d8_v8_ssse3(const unsigned char *src_ptr,
const unsigned int src_pitch,
unsigned char *output_ptr,
unsigned int out_pitch,
unsigned int output_height,
const short *filter);
void vp9_filter_block1d8_h8_ssse3(const unsigned char *src_ptr,
const unsigned int src_pitch,
unsigned char *output_ptr,
unsigned int out_pitch,
unsigned int output_height,
const short *filter);
void vp9_convolve8_horiz_ssse3(const uint8_t *src, int src_stride,
uint8_t *dst, int dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, int y_step_q4,
int w, int h) {
if (x_step_q4 == 16 && filter_x[3] != 128) {
while (w >= 16) {
vp9_filter_block1d16_h8_ssse3(src, src_stride,
dst, dst_stride,
h, filter_x);
src += 16;
dst += 16;
w -= 16;
}
while (w >= 8) {
vp9_filter_block1d8_h8_ssse3(src, src_stride,
dst, dst_stride,
h, filter_x);
src += 8;
dst += 8;
w -= 8;
}
}
if (w) {
vp9_convolve8_horiz_c(src, src_stride, dst, dst_stride,
filter_x, x_step_q4, filter_y, y_step_q4,
w, h);
}
}
void vp9_convolve8_vert_ssse3(const uint8_t *src, int src_stride,
uint8_t *dst, int dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, int y_step_q4,
int w, int h) {
if (y_step_q4 == 16 && filter_y[3] != 128) {
while (w >= 16) {
vp9_filter_block1d16_v8_ssse3(src - src_stride * 3, src_stride,
dst, dst_stride,
h, filter_y);
src += 16;
dst += 16;
w -= 16;
}
while (w >= 8) {
vp9_filter_block1d8_v8_ssse3(src - src_stride * 3, src_stride,
dst, dst_stride,
h, filter_y);
src += 8;
dst += 8;
w -= 8;
}
}
if (w) {
vp9_convolve8_vert_c(src, src_stride, dst, dst_stride,
filter_x, x_step_q4, filter_y, y_step_q4,
w, h);
}
}
void vp9_convolve8_ssse3(const uint8_t *src, int src_stride,
uint8_t *dst, int dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, int y_step_q4,
int w, int h) {
DECLARE_ALIGNED_ARRAY(16, unsigned char, fdata2, 16*23);
// check w/h due to fixed size fdata2 array
assert(w <= 16);
assert(h <= 16);
if (x_step_q4 == 16 && y_step_q4 == 16 &&
filter_x[3] != 128 && filter_y[3] != 128) {
if (w == 16) {
vp9_filter_block1d16_h8_ssse3(src - 3 * src_stride, src_stride,
fdata2, 16,
h + 7, filter_x);
vp9_filter_block1d16_v8_ssse3(fdata2, 16,
dst, dst_stride,
h, filter_y);
return;
}
if (w == 8) {
vp9_filter_block1d8_h8_ssse3(src - 3 * src_stride, src_stride,
fdata2, 16,
h + 7, filter_x);
vp9_filter_block1d8_v8_ssse3(fdata2, 16,
dst, dst_stride,
h, filter_y);
return;
}
}
vp9_convolve8_c(src, src_stride, dst, dst_stride,
filter_x, x_step_q4, filter_y, y_step_q4,
w, h);
}
#endif
;
; Copyright (c) 2010 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 "vpx_ports/x86_abi_support.asm"
;/************************************************************************************
; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
; input pixel array has output_height rows. This routine assumes that output_height is an
; even number. This function handles 8 pixels in horizontal direction, calculating ONE
; rows each iteration to take advantage of the 128 bits operations.
;
; This is an implementation of some of the SSE optimizations first seen in ffvp8
;
;*************************************************************************************/
;void vp9_filter_block1d8_v8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pitch,
; unsigned char *output_ptr,
; unsigned int out_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d8_v8_ssse3) PRIVATE
sym(vp9_filter_block1d8_v8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
movd xmm5, rcx
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rdx, DWORD PTR arg(1) ;pixels_per_line
%if ABI_IS_32BIT=0
movsxd r8, DWORD PTR arg(3) ;out_pitch
%endif
mov rax, rsi
movsxd rcx, DWORD PTR arg(4) ;output_height
add rax, rdx
lea rbx, [rdx + rdx*4]
add rbx, rdx ;pitch * 6
.vp9_filter_block1d8_v8_ssse3_loop:
movq xmm0, [rsi] ;A
movq xmm1, [rsi + rdx] ;B
movq xmm2, [rsi + rdx * 2] ;C
movq xmm3, [rax + rdx * 2] ;D
movq xmm4, [rsi + rdx * 4] ;E
movq xmm5, [rax + rdx * 4] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx] ;G
movq xmm7, [rax + rbx] ;H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
paddsw xmm0, xmm2
paddsw xmm0, krd
paddsw xmm4, xmm6
paddsw xmm0, xmm4
psraw xmm0, 7
packuswb xmm0, xmm0
add rsi, rdx
add rax, rdx
movq [rdi], xmm0
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(3) ;out_pitch
%else
add rdi, r8
%endif
dec rcx
jnz .vp9_filter_block1d8_v8_ssse3_loop
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp9_filter_block1d16_v8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pitch,
; unsigned char *output_ptr,
; unsigned int out_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d16_v8_ssse3) PRIVATE
sym(vp9_filter_block1d16_v8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
movd xmm5, rcx
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rdx, DWORD PTR arg(1) ;pixels_per_line
%if ABI_IS_32BIT=0
movsxd r8, DWORD PTR arg(3) ;out_pitch
%endif
mov rax, rsi
movsxd rcx, DWORD PTR arg(4) ;output_height
add rax, rdx
lea rbx, [rdx + rdx*4]
add rbx, rdx ;pitch * 6
.vp9_filter_block1d16_v8_ssse3_loop:
movq xmm0, [rsi] ;A
movq xmm1, [rsi + rdx] ;B
movq xmm2, [rsi + rdx * 2] ;C
movq xmm3, [rax + rdx * 2] ;D
movq xmm4, [rsi + rdx * 4] ;E
movq xmm5, [rax + rdx * 4] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx] ;G
movq xmm7, [rax + rbx] ;H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
paddsw xmm0, xmm2
paddsw xmm0, krd
paddsw xmm4, xmm6
paddsw xmm0, xmm4
psraw xmm0, 7
packuswb xmm0, xmm0
movq [rdi], xmm0
movq xmm0, [rsi + 8] ;A
movq xmm1, [rsi + rdx + 8] ;B
movq xmm2, [rsi + rdx * 2 + 8] ;C
movq xmm3, [rax + rdx * 2 + 8] ;D
movq xmm4, [rsi + rdx * 4 + 8] ;E
movq xmm5, [rax + rdx * 4 + 8] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx + 8] ;G
movq xmm7, [rax + rbx + 8] ;H
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
paddsw xmm0, xmm2
paddsw xmm4, xmm6
paddsw xmm0, krd
paddsw xmm0, xmm4
psraw xmm0, 7
packuswb xmm0, xmm0
add rsi, rdx
add rax, rdx
movq [rdi+8], xmm0
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(3) ;out_pitch
%else
add rdi, r8
%endif
dec rcx
jnz .vp9_filter_block1d16_v8_ssse3_loop
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp9_filter_block1d8_h8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
; unsigned char *output_ptr,
; unsigned int output_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d8_h8_ssse3) PRIVATE
sym(vp9_filter_block1d8_h8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
movd xmm5, rcx
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
; movdqa krd, xmm5
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
movsxd rdx, dword ptr arg(3) ;output_pitch
movsxd rcx, dword ptr arg(4) ;output_height
.filter_block1d8_h8_rowloop_ssse3:
movq xmm0, [rsi - 3] ; -3 -2 -1 0 1 2 3 4
; movq xmm3, [rsi + 4] ; 4 5 6 7 8 9 10 11
movq xmm3, [rsi + 5] ; 5 6 7 8 9 10 11 12
;note: if we create a k0_k7 filter, we can save a pshufb
; punpcklbw xmm0, xmm3 ; -3 4 -2 5 -1 6 0 7 1 8 2 9 3 10 4 11
punpcklqdq xmm0, xmm3
movdqa xmm1, xmm0
pshufb xmm0, [GLOBAL(shuf_t0t1)]
pmaddubsw xmm0, k0k1
movdqa xmm2, xmm1
pshufb xmm1, [GLOBAL(shuf_t2t3)]
pmaddubsw xmm1, k2k3
movdqa xmm4, xmm2
pshufb xmm2, [GLOBAL(shuf_t4t5)]
pmaddubsw xmm2, k4k5
pshufb xmm4, [GLOBAL(shuf_t6t7)]
pmaddubsw xmm4, k6k7
paddsw xmm0, xmm1
paddsw xmm0, xmm2