mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-27 12:40:05 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

dnn: use dispatcher for Winograd

Browse Source
This commit is contained in:
Maksim Shabunin 2024-09-17 13:28:51 +03:00
parent 5817b562b3
commit 9d64e2959f
6 changed files with 1059 additions and 1111 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
modules/dnn/CMakeLists.txt
View File

@ -8,7 +8,7 @@ ocv_add_dispatched_file_force_all("layers/layers_common" AVX AVX2 AVX512_SKX RVV
ocv_add_dispatched_file_force_all("int8layers/layers_common" AVX2 AVX512_SKX RVV LASX)
ocv_add_dispatched_file_force_all("layers/cpu_kernels/conv_block" AVX AVX2 NEON NEON_FP16)
ocv_add_dispatched_file_force_all("layers/cpu_kernels/conv_depthwise" AVX AVX2 RVV LASX)
ocv_add_dispatched_file_force_all("layers/cpu_kernels/conv_winograd_f63" AVX AVX2 NEON_FP16)
ocv_add_dispatched_file("layers/cpu_kernels/conv_winograd_f63" AVX AVX2 NEON NEON_FP16)
ocv_add_dispatched_file_force_all("layers/cpu_kernels/fast_gemm_kernels" AVX AVX2 NEON LASX)
ocv_add_module(dnn opencv_core opencv_imgproc WRAP python java objc js)

608
modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.cpp
View File

@ -12,28 +12,21 @@
#include "../../precomp.hpp"
#include "convolution.hpp"
#include "conv_winograd_f63.simd.hpp"
#include "layers/cpu_kernels/conv_winograd_f63.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
namespace cv { namespace dnn {
#if CV_NEON || CV_SIMD128 || CV_TRY_AVX2
enum { VEC_ALIGN = 32, DFT_TYPE = CV_32F }; // Memory alignment.
void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32);
/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
float* outptr, int Cg, const int winoIblock, const int winoAtomF32);
/*Output transform*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep, float* bpptr, int bpstep, float* outptr, int outstep,
float bias, float minval, float maxval, bool ifMinMaxAct);
int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr<FastConv>& conv,
int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct)
{
const cv::dnn::Winofunc func =
conv->useFP16 ? cv::dnn::getWinofunc_F16()
: (conv->useAVX || conv->useAVX2 || conv->useNEON || conv->useRVV || conv->useSIMD128) ? cv::dnn::getWinofunc_F32()
: cv::dnn::Winofunc::empty();
if (!func.isGood())
return 0;
Mat input = _input.getMat();
Mat output = _output.getMat();
Mat fusedAddMat = _fusedAddMat.getMat();
@ -52,42 +45,10 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
int ngroups = conv->ngroups, Cg = C/ngroups, Kg = K/ngroups;
const int CONV_WINO_KBLOCK = 4;
#if (CV_NEON && CV_NEON_AARCH64)
const int CONV_WINO_IBLOCK = 6;
#elif CV_TRY_AVX || CV_TRY_AVX2
const int CONV_WINO_IBLOCK = (conv->useAVX || conv->useAVX2) ? 6 : 3;
#else
const int CONV_WINO_IBLOCK = 3;
#endif
#if CV_TRY_AVX || CV_TRY_AVX2
const int CONV_WINO_ATOM_F32 = (conv->useAVX || conv->useAVX2) ? 8 : 4;
#else
const int CONV_WINO_ATOM_F32 = 4;
#endif
const int CONV_WINO_NATOMS_F32 = CONV_WINO_AREA / CONV_WINO_ATOM_F32; // for AVX2, it is 8, otherwise, it's 16.
int CONV_WINO_ATOM = CONV_WINO_ATOM_F32;
int CONV_WINO_NATOMS = CONV_WINO_NATOMS_F32;
#ifdef CONV_ARM_FP16
// FP 16
const int CONV_WINO_ATOM_F16 = CONV_WINO_ATOM_F32 * 2;
const int CONV_WINO_NATOMS_F16 = CONV_WINO_AREA / CONV_WINO_ATOM_F16;
#endif
int esz = sizeof(float );
#ifdef CONV_ARM_FP16
const bool useFP16 = conv->useFP16;
if (useFP16)
{
// works at FP 16.
CONV_WINO_ATOM = CONV_WINO_ATOM_F16;
CONV_WINO_NATOMS = CONV_WINO_NATOMS_F16;
esz = sizeof(__fp16);
}
#endif
const int CONV_WINO_IBLOCK = func.iblock;
const int CONV_WINO_ATOM = func.natom;
const int CONV_WINO_NATOMS = CONV_WINO_AREA / CONV_WINO_ATOM;
const int esz = func.esz;
int Kg_nblocks = (Kg + CONV_WINO_KBLOCK - 1)/CONV_WINO_KBLOCK;
const size_t inp_planesize = (size_t)Hi*Wi;
@ -175,35 +136,7 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
inptr = inpbuf;
inpstep = CONV_WINO_SIZE;
}
#if CV_TRY_AVX2
if (conv->useAVX2)
opt_AVX2::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
else
#endif
#if CV_TRY_AVX
if (conv->useAVX)
opt_AVX::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
else
#endif
#if CV_NEON && CV_NEON_AARCH64
if (conv->useNEON)
{
#ifdef CONV_ARM_FP16
if (useFP16)
{
opt_NEON_FP16::winofunc_BtXB_8x8_F16(inptr, inpstep, inwptr, Cg, CONV_WINO_IBLOCK,
CONV_WINO_ATOM);
}
else
#endif
opt_NEON::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK,
CONV_WINO_ATOM);
}
else
#endif
winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
func.BtXB_8x8(inptr, inpstep, (uchar*)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
}
else
{
@ -219,18 +152,20 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
// apply inverse Winograd transforms to the sums,
// add bias, apply activation function if any and store the results.
char* wptr0 = nullptr;
#ifdef CONV_ARM_FP16
if (useFP16)
if (esz == 2)
{
CV_Assert(!conv->weightsWinoBuf_FP16.empty());
wptr0 = (char *)conv->getWeightsWinoFP16();
}
else
#endif
else if (esz == 4)
{
CV_Assert(!conv->weightsWinoBuf.empty());
wptr0 = (char *)conv->getWeightsWino();
}
else
{
CV_Error(Error::StsError, "Impossible configuration");
}
parallel_for_(Range(0, ntasks), [&](const Range& r0) {
for (int task_id = r0.start; task_id < r0.end; task_id++)
@ -271,36 +206,9 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
char* inwptr = wbuf_all + inwofs * esz;
char* wptr = wptr0 + wofs * esz;
#if CV_TRY_AVX2
if (conv->useAVX2)
opt_AVX2::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
else
#endif
#if CV_TRY_AVX
if (conv->useAVX)
opt_AVX::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
else
#endif
#if CV_NEON && CV_NEON_AARCH64
if (conv->useNEON)
{
#ifdef CONV_ARM_FP16
if (useFP16)
{
opt_NEON_FP16::winofunc_accum_F16(inwptr, wptr, out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
}
else
#endif
opt_NEON::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
}
else
#endif
winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
func.accum((uchar*)inwptr, (uchar*)wptr, (uchar*)out_wbuf, Cg,
block_id1 - block_id0, CONV_WINO_IBLOCK,
CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
for (int k = k0; k < k1; k++)
{
@ -336,37 +244,10 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
dx1*sizeof(pbptr0[0]));
}
}
#if CV_TRY_AVX2
if (conv->useAVX2)
opt_AVX2::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
else
#endif
#if CV_TRY_AVX
if (conv->useAVX)
opt_AVX::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
else
#endif
#if CV_NEON && CV_NEON_AARCH64
// NEON optimization is only for ARMv8 device, and for ARMv7 device, we use the Universal intrinsics.
if (conv->useNEON)
{
#ifdef CONV_ARM_FP16
if (useFP16)
{
opt_NEON_FP16::winofunc_AtXA_8x8_F16(out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA * esz, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
}
else
#endif
opt_NEON::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
}
else
#endif
winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
const int count = ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA;
func.AtXA_8x8((uchar*)out_wbuf + count * esz, CONV_WINO_SIZE,
bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
if (partial)
{
@ -383,441 +264,4 @@ int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _outpu
return 1;
}
/****************************************************************************************\
SIMD for winograd function
\****************************************************************************************/
#if CV_SIMD128
void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32)
{
#if 1
CV_Assert(winoIblock == 3 && winoKblock == 4 && winoAtomF32 == 4);
for (int atom_id = 0; atom_id < winoNatomF32; atom_id++,
outbuf += winoAtomF32)
{
v_float32x4 s00 = v_setzero_f32(), s01 = s00, s02 = s00;
v_float32x4 s10 = v_setzero_f32(), s11 = s00, s12 = s00;
v_float32x4 s20 = v_setzero_f32(), s21 = s00, s22 = s00;
v_float32x4 s30 = v_setzero_f32(), s31 = s00, s32 = s00;
for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
wptr += winoKblock*winoAtomF32)
{
v_float32x4 x0, x1, x2;
x0 = v_load(inwptr);
x1 = v_load(inwptr + 4);
x2 = v_load(inwptr + 8);
v_float32x4 w0 = v_load(wptr);
s00 = v_fma(w0, x0, s00);
s01 = v_fma(w0, x1, s01);
s02 = v_fma(w0, x2, s02);
w0 = v_load(wptr + 4);
s10 = v_fma(w0, x0, s10);
s11 = v_fma(w0, x1, s11);
s12 = v_fma(w0, x2, s12);
w0 = v_load(wptr + 8);
s20 = v_fma(w0, x0, s20);
s21 = v_fma(w0, x1, s21);
s22 = v_fma(w0, x2, s22);
w0 = v_load(wptr + 12);
s30 = v_fma(w0, x0, s30);
s31 = v_fma(w0, x1, s31);
s32 = v_fma(w0, x2, s32);
}
v_store(outbuf, s00);
v_store(outbuf + 1*64, s01);
v_store(outbuf + 2*64, s02);
v_store(outbuf + 3*64, s10);
v_store(outbuf + 4*64, s11);
v_store(outbuf + 5*64, s12);
v_store(outbuf + 6*64, s20);
v_store(outbuf + 7*64, s21);
v_store(outbuf + 8*64, s22);
v_store(outbuf + 9*64, s30);
v_store(outbuf + 10*64, s31);
v_store(outbuf + 11*64, s32);
}
#else
// Naive C++ code, the code should never be run here.
for (int atom_id = 0; atom_id < winoNatomF32;
atom_id++, outbuf += winoAtomF32)
{
float sumbuf[winoIblock*winoKblock*winoAtomF32];
memset(sumbuf, 0, sizeof(sumbuf));
for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
wptr += winoKblock*winoAtomF32)
{
for (int i = 0; i < winoKblock; i++)
{
for (int j = 0; j < winoIblock; j++)
{
int i_ = i*winoAtomF32;
int j_ = j*winoAtomF32;
int ij_ = i_*winoIblock + j_;
float s0 = inwptr[j_ + 0]*wptr[i_ + 0];
float s1 = inwptr[j_ + 1]*wptr[i_ + 1];
float s2 = inwptr[j_ + 2]*wptr[i_ + 2];
float s3 = inwptr[j_ + 3]*wptr[i_ + 3];
sumbuf[ij_ + 0] += s0;
sumbuf[ij_ + 1] += s1;
sumbuf[ij_ + 2] += s2;
sumbuf[ij_ + 3] += s3;
}
}
}
for (int ij = 0; ij < winoKblock*winoIblock; ij++)
{
int ij_ = ij*winoAtomF32;
int ij_out = ij*CONV_WINO_AREA;
outbuf[ij_out + 0] = sumbuf[ij_ + 0];
outbuf[ij_out + 1] = sumbuf[ij_ + 1];
outbuf[ij_out + 2] = sumbuf[ij_ + 2];
outbuf[ij_out + 3] = sumbuf[ij_ + 3];
}
}
#endif
}
/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
float* outptr, int Cg, const int winoIblock, const int winoAtomF32)
{
CV_Assert(winoIblock == 3 && winoAtomF32 == 4);
v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4);
v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4);
v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4);
v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4);
v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4);
v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4);
v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4);
v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4);
v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51, z60, z61, z70, z71;
{
/* Y[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*X */
/* Y[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*X */
v_float32x4 q5_25 = v_setall_f32(5.25f), t00, t01, t10, t11;
t00 = v_sub(x40, x20);
t01 = v_sub(x41, x21);
t10 = v_sub(x30, x50);
t11 = v_sub(x31, x51);
v_float32x4 y00 = v_fma(t00, q5_25, v_sub(x00, x60));
v_float32x4 y01 = v_fma(t01, q5_25, v_sub(x01, x61));
v_float32x4 y70 = v_fma(t10, q5_25, v_sub(x70, x10));
v_float32x4 y71 = v_fma(t11, q5_25, v_sub(x71, x11));
/* Y[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*X */
/* Y[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*X */
v_float32x4 qm4_25 = v_setall_f32(-4.25f);
t00 = v_fma(x30, qm4_25, v_add(x10, x50));
t01 = v_fma(x31, qm4_25, v_add(x11, x51));
t10 = v_fma(x40, qm4_25, v_add(x20, x60));
t11 = v_fma(x41, qm4_25, v_add(x21, x61));
v_float32x4 y10 = v_add(t00, t10), y11 = v_add(t01, t11);
v_float32x4 y20 = v_sub(t10, t00), y21 = v_sub(t11, t01);
/* Y[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*X */
/* Y[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*X */
v_float32x4 q0_5 = v_setall_f32(0.5f), q0_25 = v_setall_f32(0.25f);
v_float32x4 qm2_5 = v_setall_f32(-2.5f), qm1_25 = v_setall_f32(-1.25f);
t00 = v_fma(x10, q0_5, v_add(x50, x50));
t01 = v_fma(x11, q0_5, v_add(x51, x51));
t10 = v_fma(x20, q0_25, x60);
t11 = v_fma(x21, q0_25, x61);
t00 = v_fma(x30, qm2_5, t00);
t01 = v_fma(x31, qm2_5, t01);
t10 = v_fma(x40, qm1_25, t10);
t11 = v_fma(x41, qm1_25, t11);
v_float32x4 y30 = v_add(t00, t10), y31 = v_add(t01, t11);
v_float32x4 y40 = v_sub(t10, t00), y41 = v_sub(t11, t01);
/* Y[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*X */
/* Y[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*X */
v_float32x4 q4 = v_setall_f32(4.f), qm5 = v_setall_f32(-5.f);
t00 = v_fma(x50, q0_5, v_add(x10, x10));
t01 = v_fma(x51, q0_5, v_add(x11, x11));
t10 = v_fma(x20, q4 , x60);
t11 = v_fma(x21, q4 , x61);
t00 = v_fma(x30, qm2_5, t00);
t01 = v_fma(x31, qm2_5, t01);
t10 = v_fma(x40, qm5 , t10);
t11 = v_fma(x41, qm5 , t11);
v_float32x4 y50 = v_add(t00, t10), y51 = v_add(t01, t11);
v_float32x4 y60 = v_sub(t10, t00), y61 = v_sub(t11, t01);
/* transpose 8x8 matrix with v_transpose4x4 */
v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710;
v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300);
v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310);
v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700);
v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710);
/* Z[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*Y */
/* Z[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*Y */
t00 = v_sub(y010, y200);
t01 = v_sub(y410, y600);
t10 = v_sub(y300, y110);
t11 = v_sub(y700, y510);
z00 = v_fma(t00, q5_25, v_sub(y000, y210));
z01 = v_fma(t01, q5_25, v_sub(y400, y610));
z70 = v_fma(t10, q5_25, v_sub(y310, y100));
z71 = v_fma(t11, q5_25, v_sub(y710, y500));
/* Z[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*Y */
/* Z[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*Y */
t00 = v_fma(y300, qm4_25, v_add(y100, y110));
t01 = v_fma(y700, qm4_25, v_add(y500, y510));
t10 = v_fma(y010, qm4_25, v_add(y200, y210));
t11 = v_fma(y410, qm4_25, v_add(y600, y610));
z10 = v_add(t00, t10); z11 = v_add(t01, t11);
z20 = v_sub(t10, t00); z21 = v_sub(t11, t01);
/* Z[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*Y */
/* Z[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*Y */
t00 = v_fma(y100, q0_5, v_add(y110, y110));
t01 = v_fma(y500, q0_5, v_add(y510, y510));
t10 = v_fma(y200, q0_25, y210);
t11 = v_fma(y600, q0_25, y610);
t00 = v_fma(y300, qm2_5, t00);
t01 = v_fma(y700, qm2_5, t01);
t10 = v_fma(y010, qm1_25, t10);
t11 = v_fma(y410, qm1_25, t11);
z30 = v_add(t00, t10); z31 = v_add(t01, t11);
z40 = v_sub(t10, t00); z41 = v_sub(t11, t01);
/* Z[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*Y */
/* Z[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*Y */
t00 = v_fma(y110, q0_5, v_add(y100, y100));
t01 = v_fma(y510, q0_5, v_add(y500, y500));
t10 = v_fma(y200, q4, y210);
t11 = v_fma(y600, q4, y610);
t00 = v_fma(y300, qm2_5, t00);
t01 = v_fma(y700, qm2_5, t01);
t10 = v_fma(y010, qm5, t10);
t11 = v_fma(y410, qm5, t11);
z50 = v_add(t00, t10); z51 = v_add(t01, t11);
z60 = v_sub(t10, t00); z61 = v_sub(t11, t01);
}
const int outstep = winoIblock*winoAtomF32*Cg;
v_store(outptr, z00);
v_store(outptr + outstep, z01);
v_store(outptr + outstep*2, z10);
v_store(outptr + outstep*3, z11);
v_store(outptr + outstep*4, z20);
v_store(outptr + outstep*5, z21);
v_store(outptr + outstep*6, z30);
v_store(outptr + outstep*7, z31);
v_store(outptr + outstep*8, z40);
v_store(outptr + outstep*9, z41);
v_store(outptr + outstep*10, z50);
v_store(outptr + outstep*11, z51);
v_store(outptr + outstep*12, z60);
v_store(outptr + outstep*13, z61);
v_store(outptr + outstep*14, z70);
v_store(outptr + outstep*15, z71);
}
/*Output transform*/
/* Inverse Winograd 8x8 transform:
out = (A'*inp*A)', where
inp is input 8x8 FP32 matrix,
A' is
[1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f,
0.f, 1.f, -1.f, 2.f, -2.f, 0.5f, -0.5f, 0.f,
0.f, 1.f, 1.f, 4.f, 4.f, 0.25f, 0.25f, 0.f,
0.f, 1.f, -1.f, 8.f, -8.f, 0.125f, -0.125f, 0.f,
0.f, 1.f, 1.f, 16.f, 16.f, 1.f/16, 1.f/16, 0.f,
0.f, 1.f, -1.f, 32.f, -32.f, 1.f/32, -1.f/32, 1.f]
inp is pre-loaded into xij registers,
out will be stored in zij, where (0<=i<=7 for x, 0<=i<=5 for z), 0<=j<=1.
After the inverse transform is done, we add bias,
optionally add results from the earlier tensors (by-pass),
optionally apply activation function and then
store the final results.
That is, after both forward and then inverse transformation,
we get non-transposed result.
Of course, for the correct work of Winograd-based convolution,
the Winograd-transformed weights should also be transposed.
init_conv() (see OpConv.fx) takes care of that.
*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep,
float* bpptr, int bpstep, float* outptr, int outstep,
float bias, float minval, float maxval, bool ifMinMaxAct)
{
v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4);
v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4);
v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4);
v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4);
v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4);
v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4);
v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4);
v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4);
v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51;
{
v_float32x4 s12_0, s12_1, s34_0, s34_1, s56_0, s56_1;
s12_0 = v_add(x10, x20); s12_1 = v_add(x11, x21);
s34_0 = v_add(x30, x40); s34_1 = v_add(x31, x41);
s56_0 = v_add(x50, x60); s56_1 = v_add(x51, x61);
v_float32x4 y00 = v_add(v_add(v_add(x00, s12_0), s34_0), s56_0);
v_float32x4 y01 = v_add(v_add(v_add(x01, s12_1), s34_1), s56_1);
v_float32x4 a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f);
v_float32x4 y20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
v_float32x4 y21 = v_fma(s56_1, a0 ,v_fma(s34_1, a1, s12_1) );
a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f);
v_float32x4 y40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
v_float32x4 y41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
s12_0 = v_sub(x10, x20); s12_1 = v_sub(x11, x21);
s34_0 = v_sub(x30, x40); s34_1 = v_sub(x31, x41);
s56_0 = v_sub(x50, x60); s56_1 = v_sub(x51, x61);
a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.f);
v_float32x4 y50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(x70, s12_0)));
v_float32x4 y51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(x71, s12_1)));
a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.f);
v_float32x4 y10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
v_float32x4 y11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.f);
v_float32x4 y30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
v_float32x4 y31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
v_float32x4 y60 = v_setall_f32(0.f), y61 = y60, y70 = y60, y71 = y60;
/* transpose 8x8 matrix with v_transpose4x4 */
v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710;
v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300);
v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310);
v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700);
v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710);
s12_0 = v_add(y100, y200); s12_1 = v_add(y500, y600);
s34_0 = v_add(y300, y010); s34_1 = v_add(y700, y410);
s56_0 = v_add(y110, y210); s56_1 = v_add(y510, y610);
z00 = v_add(v_add(v_add(y000, s12_0), s34_0), s56_0);
z01 = v_add(v_add(v_add(y400, s12_1), s34_1), s56_1);
a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f);
z20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
z21 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f);
z40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
z41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
s12_0 = v_sub(y100, y200); s12_1 = v_sub(y500, y600);
s34_0 = v_sub(y300, y010); s34_1 = v_sub(y700, y410);
s56_0 = v_sub(y110, y210); s56_1 = v_sub(y510, y610);
a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.0f);
z50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(y310, s12_0)));
z51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(y710, s12_1)));
a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.0f);
z10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
z11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.0f);
z30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
z31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));
v_float32x4 vbias = v_setall_f32(bias);
z00 = v_add(z00, vbias);
z01 = v_add(z01, vbias);
z10 = v_add(z10, vbias);
z11 = v_add(z11, vbias);
z20 = v_add(z20, vbias);
z21 = v_add(z21, vbias);
z30 = v_add(z30, vbias);
z31 = v_add(z31, vbias);
z40 = v_add(z40, vbias);
z41 = v_add(z41, vbias);
z50 = v_add(z50, vbias);
z51 = v_add(z51, vbias);
}
if (bpptr)
{
z00 = v_add(z00, v_load(bpptr));
z01 = v_add(z01, v_load_low(bpptr + 4));
z10 = v_add(z10, v_load(bpptr + bpstep));
z11 = v_add(z11, v_load_low(bpptr + bpstep + 4));
z20 = v_add(z20, v_load(bpptr + bpstep * 2));
z21 = v_add(z21, v_load_low(bpptr + bpstep * 2 + 4));
z30 = v_add(z30, v_load(bpptr + bpstep * 3));
z31 = v_add(z31, v_load_low(bpptr + bpstep * 3 + 4));
z40 = v_add(z40, v_load(bpptr + bpstep * 4));
z41 = v_add(z41, v_load_low(bpptr + bpstep * 4 + 4));
z50 = v_add(z50, v_load(bpptr + bpstep * 5));
z51 = v_add(z51, v_load_low(bpptr + bpstep * 5 + 4));
}
if (ifMinMaxAct)
{
v_float32x4 vmax = v_setall_f32(maxval);
v_float32x4 vmin = v_setall_f32(minval);
z00 = v_min(v_max(z00, vmin), vmax);
z01 = v_min(v_max(z01, vmin), vmax);
z10 = v_min(v_max(z10, vmin), vmax);
z11 = v_min(v_max(z11, vmin), vmax);
z20 = v_min(v_max(z20, vmin), vmax);
z21 = v_min(v_max(z21, vmin), vmax);
z30 = v_min(v_max(z30, vmin), vmax);
z31 = v_min(v_max(z31, vmin), vmax);
z40 = v_min(v_max(z40, vmin), vmax);
z41 = v_min(v_max(z41, vmin), vmax);
z50 = v_min(v_max(z50, vmin), vmax);
z51 = v_min(v_max(z51, vmin), vmax);
}
v_store(outptr, z00);
v_store_low(outptr + 4, z01);
v_store(outptr + outstep, z10);
v_store_low(outptr + outstep + 4, z11);
v_store(outptr + outstep*2, z20);
v_store_low(outptr + outstep*2 + 4, z21);
v_store(outptr + outstep*3, z30);
v_store_low(outptr + outstep*3 + 4, z31);
v_store(outptr + outstep*4, z40);
v_store_low(outptr + outstep*4 + 4, z41);
v_store(outptr + outstep*5, z50);
v_store_low(outptr + outstep*5 + 4, z51);
}
#endif
#else
int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr<FastConv>& conv,
int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct)
{
return 0;
}
#endif
}} // namespace cv::dnn

22
modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.dispatch.cpp Normal file
View File

@ -0,0 +1,22 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "convolution.hpp"
#include "conv_winograd_f63.simd.hpp"
#include "layers/cpu_kernels/conv_winograd_f63.simd_declarations.hpp"
namespace cv {
namespace dnn {
cv::dnn::Winofunc getWinofunc_F32()
{
CV_CPU_DISPATCH(getWinofunc_F32, (), CV_CPU_DISPATCH_MODES_ALL);
}
cv::dnn::Winofunc getWinofunc_F16()
{
CV_CPU_DISPATCH(getWinofunc_F16, (), CV_CPU_DISPATCH_MODES_ALL);
}
}} // namespace cv::dnn::

476
modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.neon.cpp
View File

@ -1,476 +0,0 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "../../precomp.hpp"
#include "convolution.hpp"
#include "opencv2/core/hal/intrin.hpp"
namespace cv {
namespace dnn {
// NEON code work around.
namespace opt_NEON
{
#if CV_NEON && CV_NEON_AARCH64
/* Accumulate */
void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32)
{
CV_Assert(winoIblock == 6 && winoKblock == 4 && winoAtomF32 == 4);
if (iblock > 3)
{
for (int atom_id = 0; atom_id < winoNatomF32; atom_id++,
outbuf += winoAtomF32)
{
float32x4_t s00 = vdupq_n_f32(0.f), s01 = s00, s02 = s00, s03 = s00, s04 = s00, s05 = s00;
float32x4_t s10 = vdupq_n_f32(0.f), s11 = s00, s12 = s00, s13 = s00, s14 = s00, s15 = s00;
float32x4_t s20 = vdupq_n_f32(0.f), s21 = s00, s22 = s00, s23 = s00, s24 = s00, s25 = s00;
float32x4_t s30 = vdupq_n_f32(0.f), s31 = s00, s32 = s00, s33 = s00, s34 = s00, s35 = s00;
for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
wptr += winoKblock*winoAtomF32) {
float32x4_t w0 = vld1q_f32(wptr), w1 = vld1q_f32(wptr + 4);
float32x4_t w2 = vld1q_f32(wptr + 8), w3 = vld1q_f32(wptr + 12);
float32x4_t x0, x1;
x0 = vld1q_f32(inwptr);
x1 = vld1q_f32(inwptr + 4);
s00 = vfmaq_f32(s00, w0, x0);
s01 = vfmaq_f32(s01, w0, x1);
s10 = vfmaq_f32(s10, w1, x0);
s11 = vfmaq_f32(s11, w1, x1);
s20 = vfmaq_f32(s20, w2, x0);
s21 = vfmaq_f32(s21, w2, x1);
s30 = vfmaq_f32(s30, w3, x0);
s31 = vfmaq_f32(s31, w3, x1);
x0 = vld1q_f32(inwptr + 8);
x1 = vld1q_f32(inwptr + 12);
s02 = vfmaq_f32(s02, w0, x0);
s03 = vfmaq_f32(s03, w0, x1);
s12 = vfmaq_f32(s12, w1, x0);
s13 = vfmaq_f32(s13, w1, x1);
s22 = vfmaq_f32(s22, w2, x0);
s23 = vfmaq_f32(s23, w2, x1);
s32 = vfmaq_f32(s32, w3, x0);
s33 = vfmaq_f32(s33, w3, x1);
x0 = vld1q_f32(inwptr + 16);
x1 = vld1q_f32(inwptr + 20);
s04 = vfmaq_f32(s04, w0, x0);
s05 = vfmaq_f32(s05, w0, x1);
s14 = vfmaq_f32(s14, w1, x0);
s15 = vfmaq_f32(s15, w1, x1);
s24 = vfmaq_f32(s24, w2, x0);
s25 = vfmaq_f32(s25, w2, x1);
s34 = vfmaq_f32(s34, w3, x0);
s35 = vfmaq_f32(s35, w3, x1);
}
vst1q_f32(outbuf, s00);
vst1q_f32(outbuf + 1*64, s01);
vst1q_f32(outbuf + 2*64, s02);
vst1q_f32(outbuf + 3*64, s03);
vst1q_f32(outbuf + 4*64, s04);
vst1q_f32(outbuf + 5*64, s05);
vst1q_f32(outbuf + 6*64, s10);
vst1q_f32(outbuf + 7*64, s11);
vst1q_f32(outbuf + 8*64, s12);
vst1q_f32(outbuf + 9*64, s13);
vst1q_f32(outbuf + 10*64, s14);
vst1q_f32(outbuf + 11*64, s15);
vst1q_f32(outbuf + 12*64, s20);
vst1q_f32(outbuf + 13*64, s21);
vst1q_f32(outbuf + 14*64, s22);
vst1q_f32(outbuf + 15*64, s23);
vst1q_f32(outbuf + 16*64, s24);
vst1q_f32(outbuf + 17*64, s25);
vst1q_f32(outbuf + 18*64, s30);
vst1q_f32(outbuf + 19*64, s31);
vst1q_f32(outbuf + 20*64, s32);
vst1q_f32(outbuf + 21*64, s33);
vst1q_f32(outbuf + 22*64, s34);
vst1q_f32(outbuf + 23*64, s35);
}
}
else
{
for (int atom_id = 0; atom_id < winoNatomF32; atom_id++,
outbuf += winoAtomF32)
{
float32x4_t s00 = vdupq_n_f32(0.f), s01 = s00, s02 = s00;
float32x4_t s10 = vdupq_n_f32(0.f), s11 = s00, s12 = s00;
float32x4_t s20 = vdupq_n_f32(0.f), s21 = s00, s22 = s00;
float32x4_t s30 = vdupq_n_f32(0.f), s31 = s00, s32 = s00;
for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
wptr += winoKblock*winoAtomF32) {
float32x4_t w0 = vld1q_f32(wptr), w1 = vld1q_f32(wptr + 4);
float32x4_t w2 = vld1q_f32(wptr + 8), w3 = vld1q_f32(wptr + 12);
float32x4_t x0, x1, x2;
x0 = vld1q_f32(inwptr);
x1 = vld1q_f32(inwptr + 4);
x2 = vld1q_f32(inwptr + 8);
s00 = vfmaq_f32(s00, w0, x0);
s01 = vfmaq_f32(s01, w0, x1);
s02 = vfmaq_f32(s02, w0, x2);
s10 = vfmaq_f32(s10, w1, x0);
s11 = vfmaq_f32(s11, w1, x1);
s12 = vfmaq_f32(s12, w1, x2);
s20 = vfmaq_f32(s20, w2, x0);
s21 = vfmaq_f32(s21, w2, x1);
s22 = vfmaq_f32(s22, w2, x2);
s30 = vfmaq_f32(s30, w3, x0);
s31 = vfmaq_f32(s31, w3, x1);
s32 = vfmaq_f32(s32, w3, x2);
}
vst1q_f32(outbuf, s00);
vst1q_f32(outbuf + 1*64, s01);
vst1q_f32(outbuf + 2*64, s02);
vst1q_f32(outbuf + 6*64, s10);
vst1q_f32(outbuf + 7*64, s11);
vst1q_f32(outbuf + 8*64, s12);
vst1q_f32(outbuf + 12*64, s20);
vst1q_f32(outbuf + 13*64, s21);
vst1q_f32(outbuf + 14*64, s22);
vst1q_f32(outbuf + 18*64, s30);
vst1q_f32(outbuf + 19*64, s31);
vst1q_f32(outbuf + 20*64, s32);
}
}
}
#undef T4x4
#define T4x4(a, b, c, d, tr0, tr1) \
tr0 = vtrnq_f32(a, b); \
tr1 = vtrnq_f32(c, d); \
a = vcombine_f32(vget_low_f32(tr0.val[0]), vget_low_f32(tr1.val[0])); \
b = vcombine_f32(vget_low_f32(tr0.val[1]), vget_low_f32(tr1.val[1])); \
c = vcombine_f32(vget_high_f32(tr0.val[0]), vget_high_f32(tr1.val[0])); \
d = vcombine_f32(vget_high_f32(tr0.val[1]), vget_high_f32(tr1.val[1]))
/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
float* outptr, int Cg, const int winoIblock, const int winoAtomF32)
{
float32x4_t x00 = vld1q_f32(inptr), x01 = vld1q_f32(inptr + 4);
float32x4_t x10 = vld1q_f32(inptr + inpstep), x11 = vld1q_f32(inptr + inpstep + 4);
float32x4_t x20 = vld1q_f32(inptr + inpstep*2), x21 = vld1q_f32(inptr + inpstep*2 + 4);
float32x4_t x30 = vld1q_f32(inptr + inpstep*3), x31 = vld1q_f32(inptr + inpstep*3 + 4);
float32x4_t x40 = vld1q_f32(inptr + inpstep*4), x41 = vld1q_f32(inptr + inpstep*4 + 4);
float32x4_t x50 = vld1q_f32(inptr + inpstep*5), x51 = vld1q_f32(inptr + inpstep*5 + 4);
float32x4_t x60 = vld1q_f32(inptr + inpstep*6), x61 = vld1q_f32(inptr + inpstep*6 + 4);
float32x4_t x70 = vld1q_f32(inptr + inpstep*7), x71 = vld1q_f32(inptr + inpstep*7 + 4);
float32x4_t z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51, z60, z61, z70, z71;
{
/* Y[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*X */
/* Y[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*X */
float32x4_t q5_25 = vdupq_n_f32(5.25f), t00, t01, t10, t11;
t00 = vsubq_f32(x40, x20);
t01 = vsubq_f32(x41, x21);
t10 = vsubq_f32(x30, x50);
t11 = vsubq_f32(x31, x51);
float32x4_t y00 = vfmaq_f32(vsubq_f32(x00, x60), t00, q5_25);
float32x4_t y01 = vfmaq_f32(vsubq_f32(x01, x61), t01, q5_25);
float32x4_t y70 = vfmaq_f32(vsubq_f32(x70, x10), t10, q5_25);
float32x4_t y71 = vfmaq_f32(vsubq_f32(x71, x11), t11, q5_25);
/* Y[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*X */
/* Y[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*X */
float32x4_t qm4_25 = vdupq_n_f32(-4.25f);
t00 = vfmaq_f32(vaddq_f32(x10, x50), x30, qm4_25);
t01 = vfmaq_f32(vaddq_f32(x11, x51), x31, qm4_25);
t10 = vfmaq_f32(vaddq_f32(x20, x60), x40, qm4_25);
t11 = vfmaq_f32(vaddq_f32(x21, x61), x41, qm4_25);
float32x4_t y10 = vaddq_f32(t00, t10), y11 = vaddq_f32(t01, t11);
float32x4_t y20 = vsubq_f32(t10, t00), y21 = vsubq_f32(t11, t01);
/* Y[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*X */
/* Y[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*X */
float32x4_t q0_5 = vdupq_n_f32(0.5f), q0_25 = vdupq_n_f32(0.25f);
float32x4_t qm2_5 = vdupq_n_f32(-2.5f), qm1_25 = vdupq_n_f32(-1.25f);
t00 = vfmaq_f32(vaddq_f32(x50, x50), x10, q0_5);
t01 = vfmaq_f32(vaddq_f32(x51, x51), x11, q0_5);
t10 = vfmaq_f32(x60, x20, q0_25);
t11 = vfmaq_f32(x61, x21, q0_25);
t00 = vfmaq_f32(t00, x30, qm2_5);
t01 = vfmaq_f32(t01, x31, qm2_5);
t10 = vfmaq_f32(t10, x40, qm1_25);
t11 = vfmaq_f32(t11, x41, qm1_25);
float32x4_t y30 = vaddq_f32(t00, t10), y31 = vaddq_f32(t01, t11);
float32x4_t y40 = vsubq_f32(t10, t00), y41 = vsubq_f32(t11, t01);
/* Y[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*X */
/* Y[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*X */
float32x4_t q4 = vdupq_n_f32(4.f), qm5 = vdupq_n_f32(-5.f);
t00 = vfmaq_f32(vaddq_f32(x10, x10), x50, q0_5);
t01 = vfmaq_f32(vaddq_f32(x11, x11), x51, q0_5);
t10 = vfmaq_f32(x60, x20, q4);
t11 = vfmaq_f32(x61, x21, q4);
t00 = vfmaq_f32(t00, x30, qm2_5);
t01 = vfmaq_f32(t01, x31, qm2_5);
t10 = vfmaq_f32(t10, x40, qm5);
t11 = vfmaq_f32(t11, x41, qm5);
float32x4_t y50 = vaddq_f32(t00, t10), y51 = vaddq_f32(t01, t11);
float32x4_t y60 = vsubq_f32(t10, t00), y61 = vsubq_f32(t11, t01);
/* transpose 8x8 matrix in-place with some renumeration of the elements: */
/* Y: */
/* y00 y01 */
/* y10 y11 */
/* ... */
/* y70 y71 */
/* Y': */
/* y00 y40 */
/* y10 y50 */
/* y20 y60 */
/* y30 y70 */
/* y01 y41 */
/* y11 y51 */
/* y21 y61 */
/* y31 y71 */
/* in other words, y40 <-> y01, y50 <-> y11, y60 <-> y21, y70 <-> y31 */
float32x4x2_t tr0, tr1;
T4x4(y00, y10, y20, y30, tr0, tr1);
T4x4(y01, y11, y21, y31, tr0, tr1);
T4x4(y40, y50, y60, y70, tr0, tr1);
T4x4(y41, y51, y61, y71, tr0, tr1);
/* Z[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*Y */
/* Z[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*Y */
t00 = vsubq_f32(y01, y20);
t01 = vsubq_f32(y41, y60);
t10 = vsubq_f32(y30, y11);
t11 = vsubq_f32(y70, y51);
z00 = vfmaq_f32(vsubq_f32(y00, y21), t00, q5_25);
z01 = vfmaq_f32(vsubq_f32(y40, y61), t01, q5_25);
z70 = vfmaq_f32(vsubq_f32(y31, y10), t10, q5_25);
z71 = vfmaq_f32(vsubq_f32(y71, y50), t11, q5_25);
/* Z[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*Y */
/* Z[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*Y */
t00 = vfmaq_f32(vaddq_f32(y10, y11), y30, qm4_25);
t01 = vfmaq_f32(vaddq_f32(y50, y51), y70, qm4_25);
t10 = vfmaq_f32(vaddq_f32(y20, y21), y01, qm4_25);
t11 = vfmaq_f32(vaddq_f32(y60, y61), y41, qm4_25);
z10 = vaddq_f32(t00, t10); z11 = vaddq_f32(t01, t11);
z20 = vsubq_f32(t10, t00); z21 = vsubq_f32(t11, t01);
/* Z[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*Y */
/* Z[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*Y */
t00 = vfmaq_f32(vaddq_f32(y11, y11), y10, q0_5);
t01 = vfmaq_f32(vaddq_f32(y51, y51), y50, q0_5);
t10 = vfmaq_f32(y21, y20, q0_25);
t11 = vfmaq_f32(y61, y60, q0_25);
t00 = vfmaq_f32(t00, y30, qm2_5);
t01 = vfmaq_f32(t01, y70, qm2_5);
t10 = vfmaq_f32(t10, y01, qm1_25);
t11 = vfmaq_f32(t11, y41, qm1_25);
z30 = vaddq_f32(t00, t10); z31 = vaddq_f32(t01, t11);
z40 = vsubq_f32(t10, t00); z41 = vsubq_f32(t11, t01);
/* Z[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*Y */
/* Z[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*Y */
t00 = vfmaq_f32(vaddq_f32(y10, y10), y11, q0_5);
t01 = vfmaq_f32(vaddq_f32(y50, y50), y51, q0_5);
t10 = vfmaq_f32(y21, y20, q4);
t11 = vfmaq_f32(y61, y60, q4);
t00 = vfmaq_f32(t00, y30, qm2_5);
t01 = vfmaq_f32(t01, y70, qm2_5);
t10 = vfmaq_f32(t10, y01, qm5);
t11 = vfmaq_f32(t11, y41, qm5);
z50 = vaddq_f32(t00, t10); z51 = vaddq_f32(t01, t11);
z60 = vsubq_f32(t10, t00); z61 = vsubq_f32(t11, t01);
}
const int outstep = winoIblock*winoAtomF32*Cg;
vst1q_f32(outptr, z00);
vst1q_f32(outptr + outstep, z01);
vst1q_f32(outptr + outstep*2, z10);
vst1q_f32(outptr + outstep*3, z11);
vst1q_f32(outptr + outstep*4, z20);
vst1q_f32(outptr + outstep*5, z21);
vst1q_f32(outptr + outstep*6, z30);
vst1q_f32(outptr + outstep*7, z31);
vst1q_f32(outptr + outstep*8, z40);
vst1q_f32(outptr + outstep*9, z41);
vst1q_f32(outptr + outstep*10, z50);
vst1q_f32(outptr + outstep*11, z51);
vst1q_f32(outptr + outstep*12, z60);
vst1q_f32(outptr + outstep*13, z61);
vst1q_f32(outptr + outstep*14, z70);
vst1q_f32(outptr + outstep*15, z71);
}
/*Output transform*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep,
float* bpptr, int bpstep, float* outptr, int outstep,
float bias, float minval, float maxval, bool ifMinMaxAct)
{
float32x4_t x00 = vld1q_f32(inptr), x01 = vld1q_f32(inptr + 4);
float32x4_t x10 = vld1q_f32(inptr + inpstep), x11 = vld1q_f32(inptr + inpstep + 4);
float32x4_t x20 = vld1q_f32(inptr + inpstep*2), x21 = vld1q_f32(inptr + inpstep*2 + 4);
float32x4_t x30 = vld1q_f32(inptr + inpstep*3), x31 = vld1q_f32(inptr + inpstep*3 + 4);
float32x4_t x40 = vld1q_f32(inptr + inpstep*4), x41 = vld1q_f32(inptr + inpstep*4 + 4);
float32x4_t x50 = vld1q_f32(inptr + inpstep*5), x51 = vld1q_f32(inptr + inpstep*5 + 4);
float32x4_t x60 = vld1q_f32(inptr + inpstep*6), x61 = vld1q_f32(inptr + inpstep*6 + 4);
float32x4_t x70 = vld1q_f32(inptr + inpstep*7), x71 = vld1q_f32(inptr + inpstep*7 + 4);
float32x4_t z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51;
{
float32x4_t s12_0, s12_1, s34_0, s34_1, s56_0, s56_1;
s12_0 = vaddq_f32(x10, x20); s12_1 = vaddq_f32(x11, x21);
s34_0 = vaddq_f32(x30, x40); s34_1 = vaddq_f32(x31, x41);
s56_0 = vaddq_f32(x50, x60); s56_1 = vaddq_f32(x51, x61);
float32x4_t y00 = vaddq_f32(vaddq_f32(vaddq_f32(x00, s12_0), s34_0), s56_0);
float32x4_t y01 = vaddq_f32(vaddq_f32(vaddq_f32(x01, s12_1), s34_1), s56_1);
float32x4_t y20 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 4.0f), s56_0, 0.25f);
float32x4_t y21 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 4.0f), s56_1, 0.25f);
float32x4_t y40 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 16.0f), s56_0, 1.f/16);
float32x4_t y41 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 16.0f), s56_1, 1.f/16);
s12_0 = vsubq_f32(x10, x20); s12_1 = vsubq_f32(x11, x21);
s34_0 = vsubq_f32(x30, x40); s34_1 = vsubq_f32(x31, x41);
s56_0 = vsubq_f32(x50, x60); s56_1 = vsubq_f32(x51, x61);
float32x4_t y50 = vfmaq_n_f32(vfmaq_n_f32(vaddq_f32(x70, s12_0),
s34_0, 32.f), s56_0, 1.f/32);
float32x4_t y51 = vfmaq_n_f32(vfmaq_n_f32(vaddq_f32(x71, s12_1),
s34_1, 32.f), s56_1, 1.f/32);
float32x4_t y10 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 2.0f), s56_0, 0.5f);
float32x4_t y11 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 2.0f), s56_1, 0.5f);
float32x4_t y30 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 8.0f), s56_0, 0.125f);
float32x4_t y31 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 8.0f), s56_1, 0.125f);
float32x4_t y60 = vdupq_n_f32(0.f), y61 = y60, y70 = y60, y71 = y60;
/* transpose 8x8 matrix in-place with some renumeration of the elements: */
/* Y: */
/* y00 y01 */
/* y10 y11 */
/* ... */
/* y50 y51 */
/* 0 0 */
/* 0 0 */
/* Y': */
/* y00 y40 */
/* y10 y50 */
/* y20 y60 */
/* y30 y70 */
/* y01 y41 */
/* y11 y51 */
/* y21 y61 */
/* y31 y71 */
/* in other words, y40 <-> y01, y50 <-> y11, y60 <-> y21, y70 <-> y31 */
float32x4x2_t tr0, tr1;
T4x4(y00, y10, y20, y30, tr0, tr1);
T4x4(y01, y11, y21, y31, tr0, tr1);
T4x4(y40, y50, y60, y70, tr0, tr1);
T4x4(y41, y51, y61, y71, tr0, tr1);
s12_0 = vaddq_f32(y10, y20); s12_1 = vaddq_f32(y50, y60);
s34_0 = vaddq_f32(y30, y01); s34_1 = vaddq_f32(y70, y41);
s56_0 = vaddq_f32(y11, y21); s56_1 = vaddq_f32(y51, y61);
z00 = vaddq_f32(vaddq_f32(vaddq_f32(y00, s12_0), s34_0), s56_0);
z01 = vaddq_f32(vaddq_f32(vaddq_f32(y40, s12_1), s34_1), s56_1);
z20 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 4.0f), s56_0, 0.25f);
z21 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 4.0f), s56_1, 0.25f);
z40 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 16.0f), s56_0, 1.f/16);
z41 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 16.0f), s56_1, 1.f/16);
s12_0 = vsubq_f32(y10, y20); s12_1 = vsubq_f32(y50, y60);
s34_0 = vsubq_f32(y30, y01); s34_1 = vsubq_f32(y70, y41);
s56_0 = vsubq_f32(y11, y21); s56_1 = vsubq_f32(y51, y61);
z50 = vfmaq_n_f32(vfmaq_n_f32(vaddq_f32(y31, s12_0),
s34_0, 32.f), s56_0, 1.f/32);
z51 = vfmaq_n_f32(vfmaq_n_f32(vaddq_f32(y71, s12_1),
s34_1, 32.f), s56_1, 1.f/32);
z10 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 2.0f), s56_0, 0.5f);
z11 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 2.0f), s56_1, 0.5f);
z30 = vfmaq_n_f32(vfmaq_n_f32(s12_0, s34_0, 8.0f), s56_0, 0.125f);
z31 = vfmaq_n_f32(vfmaq_n_f32(s12_1, s34_1, 8.0f), s56_1, 0.125f);
float32x4_t vbias = vdupq_n_f32(bias);
z00 = vaddq_f32(z00, vbias);
z01 = vaddq_f32(z01, vbias);
z10 = vaddq_f32(z10, vbias);
z11 = vaddq_f32(z11, vbias);
z20 = vaddq_f32(z20, vbias);
z21 = vaddq_f32(z21, vbias);
z30 = vaddq_f32(z30, vbias);
z31 = vaddq_f32(z31, vbias);
z40 = vaddq_f32(z40, vbias);
z41 = vaddq_f32(z41, vbias);
z50 = vaddq_f32(z50, vbias);
z51 = vaddq_f32(z51, vbias);
}
if (bpptr)
{
float32x2_t zhalf = vdup_n_f32(0.f);
z00 = vaddq_f32(z00, vld1q_f32(bpptr));
z01 = vaddq_f32(z01, vcombine_f32(vld1_f32(bpptr + 4), zhalf));
z10 = vaddq_f32(z10, vld1q_f32(bpptr + bpstep));
z11 = vaddq_f32(z11, vcombine_f32(vld1_f32(bpptr + bpstep + 4), zhalf));
z20 = vaddq_f32(z20, vld1q_f32(bpptr + bpstep*2));
z21 = vaddq_f32(z21, vcombine_f32(vld1_f32(bpptr + bpstep*2 + 4), zhalf));
z30 = vaddq_f32(z30, vld1q_f32(bpptr + bpstep*3));
z31 = vaddq_f32(z31, vcombine_f32(vld1_f32(bpptr + bpstep*3 + 4), zhalf));
z40 = vaddq_f32(z40, vld1q_f32(bpptr + bpstep*4));
z41 = vaddq_f32(z41, vcombine_f32(vld1_f32(bpptr + bpstep*4 + 4), zhalf));
z50 = vaddq_f32(z50, vld1q_f32(bpptr + bpstep*5));
z51 = vaddq_f32(z51, vcombine_f32(vld1_f32(bpptr + bpstep*5 + 4), zhalf));
}
if (ifMinMaxAct)
{
float32x4_t vmax = vdupq_n_f32(maxval);
float32x4_t vmin = vdupq_n_f32(minval);
z00 = vminq_f32(vmaxq_f32(z00, vmin), vmax);
z01 = vminq_f32(vmaxq_f32(z01, vmin), vmax);
z10 = vminq_f32(vmaxq_f32(z10, vmin), vmax);
z11 = vminq_f32(vmaxq_f32(z11, vmin), vmax);
z20 = vminq_f32(vmaxq_f32(z20, vmin), vmax);
z21 = vminq_f32(vmaxq_f32(z21, vmin), vmax);
z30 = vminq_f32(vmaxq_f32(z30, vmin), vmax);
z31 = vminq_f32(vmaxq_f32(z31, vmin), vmax);
z40 = vminq_f32(vmaxq_f32(z40, vmin), vmax);
z41 = vminq_f32(vmaxq_f32(z41, vmin), vmax);
z50 = vminq_f32(vmaxq_f32(z50, vmin), vmax);
z51 = vminq_f32(vmaxq_f32(z51, vmin), vmax);
}
vst1q_f32(outptr, z00);
vst1_f32(outptr + 4, vget_low_f32(z01));
vst1q_f32(outptr + outstep, z10);
vst1_f32(outptr + outstep + 4, vget_low_f32(z11));
vst1q_f32(outptr + outstep*2, z20);
vst1_f32(outptr + outstep*2 + 4, vget_low_f32(z21));
vst1q_f32(outptr + outstep*3, z30);
vst1_f32(outptr + outstep*3 + 4, vget_low_f32(z31));
vst1q_f32(outptr + outstep*4, z40);
vst1_f32(outptr + outstep*4 + 4, vget_low_f32(z41));
vst1q_f32(outptr + outstep*5, z50);
vst1_f32(outptr + outstep*5 + 4, vget_low_f32(z51));
}
#endif
}
}} // namespace

1026
modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp
View File

File diff suppressed because it is too large Load Diff

36
modules/dnn/src/layers/cpu_kernels/convolution.hpp
View File

@ -6,6 +6,7 @@
#define OPENCV_FAST_CONVOLUTION_HPP
#include "opencv2/core/hal/intrin.hpp"
#include "opencv2/dnn/all_layers.hpp"
#ifndef CONV_PRAM
#define CONV_PRAM
@ -119,25 +120,30 @@ void convBlock_F32(int np, const float* a, const float* b, float* c, int ldc, bo
void convBlockMR1_F32(int np, const float* a, const float* b, float* c, const float bias, bool init_c,
const float minval, const float maxval, bool ifMinMaxAct, const int width, const int convNR);
#if CV_NEON_AARCH64
/* Accumulate */
void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
const int winoIblock, const int winoKblock, const int winoAtom, const int winoNatom);
/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
float* outptr, int Cg, const int winoIblock, const int winoAtom);
/*Output transform*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep,
float* bpptr, int bpstep, float* outptr, int outstep,
float bias, float minval, float maxval, bool ifMinMaxAct);
#endif // CV_NEON_AARCH64
#endif // CV_NEON
} // namespace opt_NEON.
// === Function tables
struct Winofunc
{
void (*accum)(const uchar* inwptr, const uchar* wptr, uchar* outbuf, int Cg, int iblock, const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32);
void (*BtXB_8x8)(const float* inptr, int inpstep, uchar* outptr, int Cg, const int winoIblock, const int winoAtomF32);
void (*AtXA_8x8)(const uchar* inptr, int inpstep, float* bpptr, int bpstep, float* outptr, int outstep, float bias, float minval, float maxval, bool ifMinMaxAct);
int iblock;
int natom;
int esz;
bool isGood() const { return accum && BtXB_8x8 && AtXA_8x8 && iblock > 0 && natom > 0 && esz > 0; }
static Winofunc empty() { return {0, 0, 0, 0, 0, 0}; }
};
// === wrapper calls (implemented in .dispatch.cpp)
Winofunc getWinofunc_F32();
Winofunc getWinofunc_F16();
} // namespace dnn
} // namespace cv