Skip to content

Add AVX2 intra prediction for 16x16 luma and 8x8 chroma #3934

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
rbenaley wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Add AVX2 intra prediction for 16x16 luma and 8x8 chroma #3934

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions codec/common/inc/intra_pred_common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
/*!
* \copy
* Copyright (c) 2009-2013, Cisco Systems
* Copyright (c) 2026, Richard Ben Aleya
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
Expand Down Expand Up @@ -56,6 +57,10 @@ extern "C" {
//for intra-prediction ASM functions
void WelsI16x16LumaPredV_sse2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsI16x16LumaPredH_sse2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
#if defined(HAVE_AVX2)
void WelsI16x16LumaPredV_avx2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsI16x16LumaPredH_avx2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
#endif//HAVE_AVX2
#endif//X86_ASM

#if defined(HAVE_NEON)
Expand Down
44 changes: 44 additions & 0 deletions codec/common/x86/intra_pred_com.asm
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
;*!
;* \copy
;* Copyright (c) 2009-2013, Cisco Systems
;* Copyright (c) 2026, Richard Ben Aleya
;* All rights reserved.
;*
;* Redistribution and use in source and binary forms, with or without
Expand Down Expand Up @@ -115,3 +116,46 @@ WELS_EXTERN WelsI16x16LumaPredV_sse2

ret

%ifdef HAVE_AVX2
;***********************************************************************
; void WelsI16x16LumaPredV_avx2(uint8_t *pred, uint8_t *pRef, int32_t stride);
;***********************************************************************
WELS_EXTERN WelsI16x16LumaPredV_avx2
%assign push_num 0
LOAD_3_PARA
SIGN_EXTENSION r2, r2d
sub r1, r2
vbroadcasti128 ymm0, [r1]
vmovdqu [r0], ymm0
vmovdqu [r0+32], ymm0
vmovdqu [r0+64], ymm0
vmovdqu [r0+96], ymm0
vmovdqu [r0+128], ymm0
vmovdqu [r0+160], ymm0
vmovdqu [r0+192], ymm0
vmovdqu [r0+224], ymm0
vzeroupper
ret

;***********************************************************************
; void WelsI16x16LumaPredH_avx2(uint8_t *pred, uint8_t *pRef, int32_t stride);
;***********************************************************************
WELS_EXTERN WelsI16x16LumaPredH_avx2
%assign push_num 0
LOAD_3_PARA
SIGN_EXTENSION r2, r2d
dec r1
%assign h_row 0
%rep 16
%if h_row > 0
add r1, r2
%endif
vpbroadcastb xmm0, [r1]
vmovdqa [r0 + h_row * 16], xmm0
%assign h_row h_row + 1
%endrep
vzeroupper
ret

%endif ; HAVE_AVX2

7 changes: 7 additions & 0 deletions codec/encoder/core/inc/get_intra_predictor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
/*!
* \copy
* Copyright (c) 2009-2013, Cisco Systems
* Copyright (c) 2026, Richard Ben Aleya
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
Expand Down Expand Up @@ -109,6 +110,12 @@ void WelsI4x4LumaPredVR_mmx (uint8_t* pPred, uint8_t* pRef, const int32_t kiStri
void WelsI4x4LumaPredHD_mmx (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsI4x4LumaPredVL_mmx (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsI4x4LumaPredHU_mmx (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);

#if defined(HAVE_AVX2)
void WelsI16x16LumaPredDc_avx2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsI16x16LumaPredPlane_avx2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
void WelsIChromaPredV_avx2 (uint8_t* pPred, uint8_t* pRef, const int32_t kiStride);
#endif//HAVE_AVX2
#endif//X86_ASM

#if defined(HAVE_NEON)
Expand Down
10 changes: 10 additions & 0 deletions codec/encoder/core/src/get_intra_predictor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
/*!
* \copy
* Copyright (c) 2009-2013, Cisco Systems
* Copyright (c) 2026, Richard Ben Aleya
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
Expand Down Expand Up @@ -719,6 +720,15 @@ void WelsInitIntraPredFuncs (SWelsFuncPtrList* pFuncList, const uint32_t kuiCpuF
pFuncList->pfGetChromaPred[C_PRED_V] = WelsIChromaPredV_sse2;
pFuncList->pfGetChromaPred[C_PRED_P] = WelsIChromaPredPlane_sse2;
}
#if defined(HAVE_AVX2)
if (kuiCpuFlag & WELS_CPU_AVX2) {
pFuncList->pfGetLumaI16x16Pred[I16_PRED_V] = WelsI16x16LumaPredV_avx2;
pFuncList->pfGetLumaI16x16Pred[I16_PRED_H] = WelsI16x16LumaPredH_avx2;
pFuncList->pfGetLumaI16x16Pred[I16_PRED_DC] = WelsI16x16LumaPredDc_avx2;
pFuncList->pfGetLumaI16x16Pred[I16_PRED_P] = WelsI16x16LumaPredPlane_avx2;
pFuncList->pfGetChromaPred[C_PRED_V] = WelsIChromaPredV_avx2;
}
#endif//HAVE_AVX2
#endif

#if defined(HAVE_MMI)
Expand Down
217 changes: 217 additions & 0 deletions codec/encoder/core/x86/intra_pred.asm
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
;*!
;* \copy
;* Copyright (c) 2009-2013, Cisco Systems
;* Copyright (c) 2026, Richard Ben Aleya
;* All rights reserved.
;*
;* Redistribution and use in source and binary forms, with or without
Expand Down Expand Up @@ -1127,3 +1128,219 @@ WELS_EXTERN WelsI16x16LumaPredDc_sse2
pop r4
pop r3
ret

%ifdef HAVE_AVX2

;***********************************************************************
; void WelsI16x16LumaPredDc_avx2(uint8_t *pred, uint8_t *pRef, int32_t stride);
;***********************************************************************
WELS_EXTERN WelsI16x16LumaPredDc_avx2
push r3
push r4
%assign push_num 2
LOAD_3_PARA
SIGN_EXTENSION r2, r2d
sub r1, r2
vmovdqa xmm0, [r1]
vpxor xmm1, xmm1, xmm1
vpsadbw xmm0, xmm0, xmm1
vpsrldq xmm1, xmm0, 8
vpaddw xmm0, xmm0, xmm1

movzx r3, byte [r1+r2-0x01]
movzx r4, byte [r1+2*r2-0x01]
add r3, r4
lea r1, [r1+r2]
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
LOAD_2_LEFT_AND_ADD
add r3, 0x10
vmovd xmm1, r3d
vpaddw xmm0, xmm0, xmm1
vpsrld xmm0, xmm0, 5
vpbroadcastb ymm0, xmm0

vmovdqu [r0], ymm0
vmovdqu [r0+32], ymm0
vmovdqu [r0+64], ymm0
vmovdqu [r0+96], ymm0
vmovdqu [r0+128], ymm0
vmovdqu [r0+160], ymm0
vmovdqu [r0+192], ymm0
vmovdqu [r0+224], ymm0
vzeroupper
pop r4
pop r3
ret

;***********************************************************************
; void WelsI16x16LumaPredPlane_avx2(uint8_t *pred, uint8_t *pRef, int32_t stride);
;***********************************************************************
WELS_EXTERN WelsI16x16LumaPredPlane_avx2
push r3
push r4
%assign push_num 2
INIT_X86_32_PIC r5
LOAD_3_PARA
PUSH_XMM 6
SIGN_EXTENSION r2, r2d
sub r1, 1
sub r1, r2

; Compute H (horizontal gradient of top row)
vpxor xmm7, xmm7, xmm7
vmovq xmm0, [r1]
vmovdqa xmm5, [pic(sse2_plane_dec)]
vpunpcklbw xmm0, xmm0, xmm7
vpmullw xmm0, xmm0, xmm5
vmovq xmm1, [r1 + 9]
vmovdqa xmm6, [pic(sse2_plane_inc)]
vpunpcklbw xmm1, xmm1, xmm7
vpmullw xmm1, xmm1, xmm6
vpsubw xmm1, xmm1, xmm0

; Horizontal sum -> H
vpsrldq xmm0, xmm1, 8
vpaddw xmm1, xmm1, xmm0
vpsrldq xmm0, xmm1, 4
vpaddw xmm1, xmm1, xmm0
vpsrldq xmm0, xmm1, 2
vpaddw xmm1, xmm1, xmm0

vmovd r3d, xmm1
movsx r3, r3w
imul r3, 5
add r3, 32
sar r3, 6 ; b = (5 * H + 32) >> 6
vmovd xmm1, r3d
vpbroadcastw ymm1, xmm1 ; ymm1 = b (16 words)

; Load top[15]
movzx r4, BYTE [r1+16]

; Load upper column (rows 0-7 of left border) using VEX instructions
sub r1, 3
vmovd xmm0, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm0, xmm0, xmm2
lea r1, [r1+2*r2]
vmovd xmm3, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm3, xmm3, xmm2
vpunpcklwd xmm0, xmm0, xmm3
lea r1, [r1+2*r2]
vmovd xmm4, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm4, xmm4, xmm2
lea r1, [r1+2*r2]
vmovd xmm3, [r1]
vmovd xmm2, [r1+r2]
lea r1, [r1+2*r2]
vpunpcklbw xmm3, xmm3, xmm2
vpunpcklwd xmm4, xmm4, xmm3
vpunpckhdq xmm0, xmm0, xmm4

; Get left[15*stride]
add r1, 3
movzx r3, BYTE [r1+8*r2]
add r4, r3
shl r4, 4 ; a = (left[15*stride] + top[15]) << 4

; Load lower column (rows 8-15 of left border) using VEX instructions
sub r1, 3
add r1, r2
vmovd xmm7, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm7, xmm7, xmm2
lea r1, [r1+2*r2]
vmovd xmm3, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm3, xmm3, xmm2
vpunpcklwd xmm7, xmm7, xmm3
lea r1, [r1+2*r2]
vmovd xmm4, [r1]
vmovd xmm2, [r1+r2]
vpunpcklbw xmm4, xmm4, xmm2
lea r1, [r1+2*r2]
vmovd xmm3, [r1]
vmovd xmm2, [r1+r2]
lea r1, [r1+2*r2]
vpunpcklbw xmm3, xmm3, xmm2
vpunpcklwd xmm4, xmm4, xmm3
vpunpckhdq xmm7, xmm7, xmm4

; Compute V gradient
vpxor xmm4, xmm4, xmm4
vpunpckhbw xmm0, xmm0, xmm4
vpmullw xmm0, xmm0, xmm5
vpunpckhbw xmm7, xmm7, xmm4
vpmullw xmm7, xmm7, xmm6
vpsubw xmm7, xmm7, xmm0

; Horizontal sum -> V
vpsrldq xmm0, xmm7, 8
vpaddw xmm7, xmm7, xmm0
vpsrldq xmm0, xmm7, 4
vpaddw xmm7, xmm7, xmm0
vpsrldq xmm0, xmm7, 2
vpaddw xmm7, xmm7, xmm0

vmovd r3d, xmm7
movsx r3, r3w
imul r3, 5
add r3, 32
sar r3, 6 ; c = (5 * V + 32) >> 6
vmovd xmm4, r3d
vpbroadcastw ymm4, xmm4 ; ymm4 = c (16 words)

; s = a + 16 + (-7)*c
add r4, 16
imul r3, -7
add r3, r4
vmovd xmm0, r3d
vpbroadcastw ymm0, xmm0 ; ymm0 = s (16 words)

; Build 256-bit multiplier: [-7,-6,-5,-4,-3,-2,-1,0 | 1,2,3,4,5,6,7,8]
vmovdqa xmm5, [pic(sse2_plane_inc_minus)]
vinserti128 ymm5, ymm5, [pic(sse2_plane_inc)], 1

xor r3, r3
.loop_plane_avx2:
vpmullw ymm2, ymm1, ymm5 ; b * [-7..8]
vpaddw ymm2, ymm2, ymm0 ; + s
vpsraw ymm2, ymm2, 5 ; >> 5
vpackuswb ymm2, ymm2, ymm2 ; pack to bytes (within each lane)
vpermq ymm2, ymm2, 0x08 ; gather low qwords from each lane
vmovdqu [r0], xmm2 ; store 16 bytes
vpaddw ymm0, ymm0, ymm4 ; s += c
add r0, 16
inc r3
cmp r3, 16
jnz .loop_plane_avx2

vzeroupper
POP_XMM
DEINIT_X86_32_PIC
pop r4
pop r3
ret

;***********************************************************************
; void WelsIChromaPredV_avx2(uint8_t *pred, uint8_t *pRef, int32_t stride);
;***********************************************************************
WELS_EXTERN WelsIChromaPredV_avx2
%assign push_num 0
LOAD_3_PARA
SIGN_EXTENSION r2, r2d
sub r1, r2
vpbroadcastq ymm0, [r1]
vmovdqu [r0], ymm0
vmovdqu [r0+32], ymm0
vzeroupper
ret

%endif ; HAVE_AVX2