OSSForks/opencv
1
0
Fork 0
mirror of https://github.com/zebrajr/opencv.git synced 2025-12-06 12:19:50 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #27723 from fanchenkong1:fastconv-wasm-scalar-opt

Browse Source 
DNN: Add large scalar fastconv kernel for WASM/no-SIMD build
This commit is contained in:
Alexander Smorkalov 2025-10-31 22:50:41 +03:00 committed by GitHub
commit 01f9bdae4c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 88 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

88
modules/dnn/src/layers/cpu_kernels/convolution.cpp
View File

@ -2048,6 +2048,85 @@ static inline void convBlock4x4(int np, const float* a, const float* b, float* c
} }
#endif #endif
#if defined(__EMSCRIPTEN__)
template<int RowBase, int NR = 24>
static inline void conv2xNR(
int np,
const float* a,
const float* b,
float* c, int ldc,
bool init_c)
{
static_assert(NR % 8 == 0, "NR must be a multiple of 8");
float* d0 = c + (RowBase + 0) * ldc;
float* d1 = c + (RowBase + 1) * ldc;
// p = 0
{
const float* bp = b;
const float a0 = a[RowBase + 0];
const float a1 = a[RowBase + 1];
if (init_c)
{
for (int j = 0; j < NR; j += 8)
{
const float b0=bp[j+0], b1=bp[j+1], b2=bp[j+2], b3=bp[j+3];
const float b4=bp[j+4], b5=bp[j+5], b6=bp[j+6], b7=bp[j+7];
d0[j+0] = b0*a0; d0[j+1] = b1*a0; d0[j+2] = b2*a0; d0[j+3] = b3*a0;
d0[j+4] = b4*a0; d0[j+5] = b5*a0; d0[j+6] = b6*a0; d0[j+7] = b7*a0;
d1[j+0] = b0*a1; d1[j+1] = b1*a1; d1[j+2] = b2*a1; d1[j+3] = b3*a1;
d1[j+4] = b4*a1; d1[j+5] = b5*a1; d1[j+6] = b6*a1; d1[j+7] = b7*a1;
}
} else
{
for (int j = 0; j < NR; j += 8)
{
const float b0=bp[j+0], b1=bp[j+1], b2=bp[j+2], b3=bp[j+3];
const float b4=bp[j+4], b5=bp[j+5], b6=bp[j+6], b7=bp[j+7];
d0[j+0] += b0*a0; d0[j+1] += b1*a0; d0[j+2] += b2*a0; d0[j+3] += b3*a0;
d0[j+4] += b4*a0; d0[j+5] += b5*a0; d0[j+6] += b6*a0; d0[j+7] += b7*a0;
d1[j+0] += b0*a1; d1[j+1] += b1*a1; d1[j+2] += b2*a1; d1[j+3] += b3*a1;
d1[j+4] += b4*a1; d1[j+5] += b5*a1; d1[j+6] += b6*a1; d1[j+7] += b7*a1;
}
}
}
// p = 1..np-1
for (int p = 1; p < np; ++p)
{
const float* bp = b + p * NR;
const int aoff = p * 4 + RowBase;
const float a0 = a[aoff + 0];
const float a1 = a[aoff + 1];
for (int j = 0; j < NR; j += 8)
{
const float b0=bp[j+0], b1=bp[j+1], b2=bp[j+2], b3=bp[j+3];
const float b4=bp[j+4], b5=bp[j+5], b6=bp[j+6], b7=bp[j+7];
d0[j+0] += b0*a0; d0[j+1] += b1*a0; d0[j+2] += b2*a0; d0[j+3] += b3*a0;
d0[j+4] += b4*a0; d0[j+5] += b5*a0; d0[j+6] += b6*a0; d0[j+7] += b7*a0;
d1[j+0] += b0*a1; d1[j+1] += b1*a1; d1[j+2] += b2*a1; d1[j+3] += b3*a1;
d1[j+4] += b4*a1; d1[j+5] += b5*a1; d1[j+6] += b6*a1; d1[j+7] += b7*a1;
}
}
}
// MR == 4, outLen == 24, scalar (no SIMD128)
static inline void convBlockNoSIMD4x24(
int np,
const float* a,
const float* b,
float* c, int ldc,
bool init_c,
int convNR)
{
CV_Assert(np > 0 && convNR == 24);
conv2xNR<0, 24>(np, a, b, c, ldc, init_c); // rows 0 & 1
conv2xNR<2, 24>(np, a, b, c, ldc, init_c); // rows 2 & 3
}
#endif
static inline void convBlockNoSIMD(int np, const float* a, const float* b, float* c, int ldc, bool init_c, const int outLen, static inline void convBlockNoSIMD(int np, const float* a, const float* b, float* c, int ldc, bool init_c, const int outLen,
const int convMR, const int convNR) const int convMR, const int convNR)
{ {
@ -2107,8 +2186,17 @@ void convBlock_F32(int np, const float* a, const float* b, float* c, int ldc, bo
return; return;
} }
convBlockNoSIMD(np, a, b, c, ldc, init_c, outLen, convMR, convNR); convBlockNoSIMD(np, a, b, c, ldc, init_c, outLen, convMR, convNR);
#elif defined(__EMSCRIPTEN__)
CV_Assert(convMR == 4);
if (outLen == 24 && convNR == 24)
{
convBlockNoSIMD4x24(np, a, b, c, ldc, init_c, convNR);
return;
}
convBlockNoSIMD(np, a, b, c, ldc, init_c, outLen, convMR, convNR);
#else #else
convBlockNoSIMD(np, a, b, c, ldc, init_c, outLen, convMR, convNR); convBlockNoSIMD(np, a, b, c, ldc, init_c, outLen, convMR, convNR);
return;
#endif #endif
} }