Skip to content
This repository has been archived by the owner on Jul 29, 2024. It is now read-only.

AVX512_changes #478

Open
wants to merge 13 commits into
base: master
Choose a base branch
from
Open

AVX512_changes #478

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
2c2c7e4
AVX512_changes
Feb 19, 2020
157cbe8
AVX512_changes_P2
deeptiag1 Feb 19, 2020
4093056
AVX512_changes_P3
deeptiag1 Feb 19, 2020
b00574f
AVX512_changes_P4
deeptiag1 Feb 20, 2020
b827613
AVX512_changes_P4
deeptiag1 Feb 20, 2020
d14b735
AVX512_changes_P4
deeptiag1 Feb 21, 2020
0bc4093
AVX512_changes_P4
deeptiag1 Feb 21, 2020
67ec8bf
AVX512_changes_Revert
deeptiag1 Feb 21, 2020
d92a7c4
changes_incorpoarted
deeptiag1 Mar 12, 2020
d983a3b
avx512-changes_v3
deeptiag1 Mar 12, 2020
06ec13d
avx512-changes_v3
deeptiag1 Apr 1, 2020
0f5ce17
avx512-changes_v3
deeptiag1 Apr 1, 2020
a23033f
avx512-changes_v3
deeptiag1 Apr 1, 2020
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
6 changes: 5 additions & 1 deletion Source/Lib/ASM_AVX2/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,10 @@ link_directories(${PROJECT_SOURCE_DIR}/Source/Lib/ASM_SSSE3/)

set(flags_to_test
-mavx2
-static-intel
-mavx512bw
-mavx512vnni
-mavx512vl
-static-intel
/Qwd10148
/Qwd10010
/Qwd10157)
Expand Down Expand Up @@ -79,6 +82,7 @@ set(ASM_AVX2_SOURCE
EbNoiseExtractAVX2.c
EbPackUnPack_Intrinsic_AVX2.c
EbPictureOperators_Intrinsic_AVX2.c
EbTransforms_Intrinsic_AVX512.c
EbTransforms_Intrinsic_AVX2.c)

if(COMPILE_AS_CPP)
Expand Down
25 changes: 14 additions & 11 deletions Source/Lib/ASM_AVX2/EbComputeSAD_SadLoopKernel_AVX512.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,6 +209,8 @@ void SadLoopKernel_AVX512_HmeL0_INTRIN(

case 16:
{
__m512i x = _mm512_setr_epi64(0x0000000000000000, 0x0001000100010001, 0x0004000400040004, 0x0005000500050005, 0x0001000100010001, 0x0002000200020002, 0x0005000500050005, 0x0006000600060006);
__m512i x1 = _mm512_setr_epi64(0x0001000100010001, 0x0002000200020002, 0x0005000500050005, 0x0006000600060006, 0x0002000200020002, 0x0003000300030003, 0x0006000600060006, 0x0007000700070007);
if (height <= 16 && searchAreaWidth <= 128)
{
for (i = 0; i<searchAreaHeight; i++)
Expand Down Expand Up @@ -243,17 +245,18 @@ void SadLoopKernel_AVX512_HmeL0_INTRIN(

for (j = 0, pRef = pRef1; j < n; j++, pRef += 16)
{
__m256i ss0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_mm_loadu_si128((__m128i*)(pRef ))), _mm_loadu_si128((__m128i*)(pRef + refStride )), 0x1);
__m256i ss1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_mm_loadu_si128((__m128i*)(pRef + 8 ))), _mm_loadu_si128((__m128i*)(pRef + refStride + 8 )), 0x1);
ss2 = _mm256_inserti128_si256(_mm256_castsi128_si256(_mm_loadu_si128((__m128i*)(pRef + 16))), _mm_loadu_si128((__m128i*)(pRef + refStride + 16)), 0x1);

__m512i ss1ftemp = _mm512_inserti64x4(_mm512_castsi256_si512(ss0), ss1, 0x1);
__m512i ss2ftemp = _mm512_inserti64x4(_mm512_castsi256_si512(ss1), ss2, 0x1);

ss3sum1_aaray[j] = _mm512_adds_epu16(ss3sum1_aaray[j], _mm512_dbsad_epu8(ref1ftemp, ss1ftemp, 0x94));
ss3sum1_aaray[j] = _mm512_adds_epu16(ss3sum1_aaray[j], _mm512_dbsad_epu8(ref2ftemp, ss1ftemp, 0xE9));
ss7sum1_aaray[j] = _mm512_adds_epu16(ss7sum1_aaray[j], _mm512_dbsad_epu8(ref3ftemp, ss2ftemp, 0x94));
ss7sum1_aaray[j] = _mm512_adds_epu16(ss7sum1_aaray[j], _mm512_dbsad_epu8(ref4ftemp, ss2ftemp, 0xE9));
__m512i ss0 = _mm512_inserti64x4(_mm512_castsi256_si512(_mm256_loadu_si256((__m256i*)(pRef))), _mm256_loadu_si256((__m256i*)(pRef + refStride )), 0x1);
__m512i ss1ftemp = _mm512_permutexvar_epi64(x, ss0);
__m512i ss2ftemp = _mm512_permutexvar_epi64(x1, ss0);
__m512i temp = ss3sum1_aaray[j];
__m512i temp1 = ss7sum1_aaray[j];

temp = _mm512_adds_epu16(temp, _mm512_dbsad_epu8(ref1ftemp, ss1ftemp, 0x94));
temp = _mm512_adds_epu16(temp, _mm512_dbsad_epu8(ref2ftemp, ss1ftemp, 0xE9));
temp1 = _mm512_adds_epu16(temp1, _mm512_dbsad_epu8(ref3ftemp, ss2ftemp, 0x94));
temp1 = _mm512_adds_epu16(temp1, _mm512_dbsad_epu8(ref4ftemp, ss2ftemp, 0xE9));
ss3sum1_aaray[j] = temp;
ss7sum1_aaray[j] = temp1;

}

Expand Down
8 changes: 8 additions & 0 deletions Source/Lib/ASM_AVX2/EbTransforms_AVX2.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,14 @@
extern "C" {
#endif

#ifdef NON_AVX512_SUPPORT
#define EbHevcTransform32_INTRIN EbHevcTransform32_AVX2_INTRIN
#else
#define EbHevcTransform32_INTRIN EbHevcTransform32_AVX512_INTRIN
#endif



void QuantizeInvQuantize8x8_AVX2_INTRIN(
EB_S16 *coeff,
const EB_U32 coeffStride,
Expand Down
134 changes: 76 additions & 58 deletions Source/Lib/ASM_AVX2/EbTransforms_Intrinsic_AVX2.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -623,7 +623,7 @@ void EbHevcTransform32_AVX2_INTRIN(EB_S16 *src, EB_U32 src_stride, EB_S16 *dst,

for (i = 0; i < 16; i++)
{
__m256i x0, x1, x2, x3;
__m256i x0, x1, x2, x3,sox0,sox5,soxa,soxf,s1x0,s1x5,s1xa,s1xf;
__m256i y0, y1, y2, y3;
__m256i a0, a1, a2, a3, a4, a5, a6, a7;
__m256i b0, b1, b2, b3, b4, b5, b6, b7;
Expand Down Expand Up @@ -652,61 +652,79 @@ void EbHevcTransform32_AVX2_INTRIN(EB_S16 *src, EB_U32 src_stride, EB_S16 *dst,
x2 = y2;
x3 = y3;

a0 = _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0x00), coeff32[0]);
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0x55), coeff32[2]));
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0xaa), coeff32[4]));
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0xff), coeff32[6]));

a1 = _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0x00), coeff32[1]);
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0x55), coeff32[3]));
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0xaa), coeff32[5]));
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(_mm256_shuffle_epi32(x0, 0xff), coeff32[7]));

a2 = _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0x00), coeff32[8]);
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0x55), coeff32[10]));
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0xaa), coeff32[12]));
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0xff), coeff32[14]));

a3 = _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0x00), coeff32[9]);
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0x55), coeff32[11]));
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0xaa), coeff32[13]));
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(_mm256_shuffle_epi32(x1, 0xff), coeff32[15]));

a4 = _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x00), coeff32[16]);
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x55), coeff32[20]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xaa), coeff32[24]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xff), coeff32[28]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x00), coeff32[32]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x55), coeff32[36]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xaa), coeff32[40]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xff), coeff32[44]));

a5 = _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x00), coeff32[17]);
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x55), coeff32[21]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xaa), coeff32[25]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xff), coeff32[29]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x00), coeff32[33]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x55), coeff32[37]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xaa), coeff32[41]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xff), coeff32[45]));

a6 = _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x00), coeff32[18]);
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x55), coeff32[22]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xaa), coeff32[26]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xff), coeff32[30]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x00), coeff32[34]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x55), coeff32[38]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xaa), coeff32[42]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xff), coeff32[46]));

a7 = _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x00), coeff32[19]);
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0x55), coeff32[23]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xaa), coeff32[27]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x2, 0xff), coeff32[31]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x00), coeff32[35]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0x55), coeff32[39]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xaa), coeff32[43]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(_mm256_shuffle_epi32(x3, 0xff), coeff32[47]));
sox0 = _mm256_shuffle_epi32(x0, 0x00);
sox5 = _mm256_shuffle_epi32(x0, 0x55);
soxa = _mm256_shuffle_epi32(x0, 0xaa);
soxf = _mm256_shuffle_epi32(x0, 0xff);
s1x0 = _mm256_shuffle_epi32(x1, 0x00);
s1x5 = _mm256_shuffle_epi32(x1, 0x55);
s1xa = _mm256_shuffle_epi32(x1, 0xaa);
s1xf = _mm256_shuffle_epi32(x1, 0xff);

a0 = _mm256_madd_epi16(sox0, coeff32[0]);
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(sox5, coeff32[2]));
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(soxa, coeff32[4]));
a0 = _mm256_add_epi32(a0, _mm256_madd_epi16(soxf, coeff32[6]));

a1 = _mm256_madd_epi16(sox0, coeff32[1]);
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(sox5, coeff32[3]));
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(soxa, coeff32[5]));
a1 = _mm256_add_epi32(a1, _mm256_madd_epi16(soxf, coeff32[7]));

a2 = _mm256_madd_epi16(s1x0, coeff32[8]);
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(s1x5, coeff32[10]));
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(s1xa, coeff32[12]));
a2 = _mm256_add_epi32(a2, _mm256_madd_epi16(s1xf, coeff32[14]));

a3 = _mm256_madd_epi16(s1x0, coeff32[9]);
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(s1x5, coeff32[11]));
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(s1xa, coeff32[13]));
a3 = _mm256_add_epi32(a3, _mm256_madd_epi16(s1xf, coeff32[15]));

sox0 = _mm256_shuffle_epi32(x2, 0x00);
sox5 = _mm256_shuffle_epi32(x2, 0x55);
soxa = _mm256_shuffle_epi32(x2, 0xaa);
soxf = _mm256_shuffle_epi32(x2, 0xff);
s1x0 = _mm256_shuffle_epi32(x3, 0x00);
s1x5 = _mm256_shuffle_epi32(x3, 0x55);
s1xa = _mm256_shuffle_epi32(x3, 0xaa);
s1xf = _mm256_shuffle_epi32(x3, 0xff);

a4 = _mm256_madd_epi16(sox0, coeff32[16]);
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(sox5, coeff32[20]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(soxa, coeff32[24]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(soxf, coeff32[28]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(s1x0, coeff32[32]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(s1x5, coeff32[36]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(s1xa, coeff32[40]));
a4 = _mm256_add_epi32(a4, _mm256_madd_epi16(s1xf, coeff32[44]));

a5 = _mm256_madd_epi16(sox0, coeff32[17]);
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(sox5, coeff32[21]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(soxa, coeff32[25]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(soxf, coeff32[29]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(s1x0, coeff32[33]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(s1x5, coeff32[37]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(s1xa, coeff32[41]));
a5 = _mm256_add_epi32(a5, _mm256_madd_epi16(s1xf, coeff32[45]));

a6 = _mm256_madd_epi16(sox0, coeff32[18]);
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(sox5, coeff32[22]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(soxa, coeff32[26]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(soxf, coeff32[30]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(s1x0, coeff32[34]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(s1x5, coeff32[38]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(s1xa, coeff32[42]));
a6 = _mm256_add_epi32(a6, _mm256_madd_epi16(s1xf, coeff32[46]));

a7 = _mm256_madd_epi16(sox0, coeff32[19]);
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(sox5, coeff32[23]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(soxa, coeff32[27]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(soxf, coeff32[31]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(s1x0, coeff32[35]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(s1x5, coeff32[39]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(s1xa, coeff32[43]));
a7 = _mm256_add_epi32(a7, _mm256_madd_epi16(s1xf, coeff32[47]));

b0 = _mm256_sra_epi32(_mm256_add_epi32(a0, o0), s0);
b1 = _mm256_sra_epi32(_mm256_add_epi32(a1, o0), s0);
Expand Down Expand Up @@ -1468,9 +1486,9 @@ EB_EXTERN void lowPrecisionTransform16x16_AVX2_INTRIN(EB_S16 *src, EB_U32 src_st
// forward 32x32 transform
EB_EXTERN void lowPrecisionTransform32x32_AVX2_INTRIN(EB_S16 *src, EB_U32 src_stride, EB_S16 *dst, EB_U32 dst_stride, EB_S16 *intermediate, EB_U32 addshift)
{
EbHevcTransform32_AVX2_INTRIN(src, src_stride, intermediate, 32, 6 + addshift);
EbHevcTransform32_INTRIN(src, src_stride, intermediate, 32, 6 + addshift);
EbHevcTranspose32_AVX2_INTRIN(intermediate, 32, dst, dst_stride);
EbHevcTransform32_AVX2_INTRIN(dst, dst_stride, intermediate, 32, 9);
EbHevcTransform32_INTRIN(dst, dst_stride, intermediate, 32, 9);
EbHevcTranspose32_AVX2_INTRIN(intermediate, 32, dst, dst_stride);
}

Expand Down
Loading