Skip to content

Improve SVE2 SynetInnerProduct16bGemmNN with svbfdot_f32 #823

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

Merged
ermig1979 merged 1 commit into from
Aug 5, 2026
Merged

Improve SVE2 SynetInnerProduct16bGemmNN with svbfdot_f32 #823

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
4 changes: 4 additions & 0 deletions docs/2026.html
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,10 @@ <h5>New features</h5>
<li>SVE2 optimizations of class SynetInnerProduct16bGemmNN.</li>
<li>NEON optimizations of class SynetInnerProduct16bGemmNN.</li>
</ul>
<h5>Improving</h5>
<ul>
<li>SVE2 optimizations of class SynetInnerProduct16bGemmNN (use of svbfdot_f32).</li>
</ul>
<h5>Renaming</h5>
<ul>
<li>Rename function SimdVectorProduct to SimdCreateMask.</li>
Expand Down
100 changes: 37 additions & 63 deletions src/Simd/SimdSve2SynetInnerProduct16bGemmNN.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,9 +49,14 @@ namespace Simd
return svlsr_n_u32_x(mask, svadd_u32_x(mask, bits, round), Base::Bf16::SHIFT);
}

SIMD_INLINE svfloat32_t BroadcastBf16(uint16_t value)
SIMD_INLINE svbfloat16_t BroadcastBf16x2(const uint16_t* src)
{
return svreinterpret_f32_u32(svdup_n_u32(uint32_t(value) << Base::Bf16::SHIFT));
return svreinterpret_bf16_u32(svdup_n_u32(uint32_t(src[0]) | (uint32_t(src[1]) << 16)));
}

SIMD_INLINE svbfloat16_t LoadBf16x2(const uint16_t* src, const svbool_t& mask)
{
return svreinterpret_bf16_u32(svld1_u32(mask, (const uint32_t*)src));
}

//-----------------------------------------------------------------------------------------
Expand Down Expand Up @@ -373,8 +378,8 @@ namespace Simd
{
const size_t F = a.F, DF = F * 2;
const svbool_t body = svptrue_b32();
svfloat32_t c00, c01, c10, c11, c20, c21, c30, c31, c40, c41, a0, b00, b01, b10, b11;
svuint32_t maskBits = svdup_n_u32(Base::Bf16::MASK);
svfloat32_t c00, c01, c10, c11, c20, c21, c30, c31, c40, c41;
svbfloat16_t a0, b0, b1;
size_t dC = a.cN, dA = a.aK, dD = p.N * a.eC;
const uint16_t* B1 = B0 + a.bK * F;
const uint16_t* A1 = A0 + 1 * dA;
Expand All @@ -401,56 +406,37 @@ namespace Simd
}
for (size_t k = 0; k < K; k += 2)
{
svuint32_t b0u = svld1_u32(body, (const uint32_t*)B0);
b00 = svreinterpret_f32_u32(svlsl_n_u32_x(body, b0u, Base::Bf16::SHIFT));
b01 = svreinterpret_f32_u32(svand_u32_x(body, b0u, maskBits));
svuint32_t b1u = svld1_u32(body, (const uint32_t*)B1);
b10 = svreinterpret_f32_u32(svlsl_n_u32_x(body, b1u, Base::Bf16::SHIFT));
b11 = svreinterpret_f32_u32(svand_u32_x(body, b1u, maskBits));
b0 = LoadBf16x2(B0, body);
b1 = LoadBf16x2(B1, body);
if (M > 0)
{
a0 = BroadcastBf16(A0[k + 0]);
c00 = svmla_f32_x(body, c00, a0, b00);
c01 = svmla_f32_x(body, c01, a0, b10);
a0 = BroadcastBf16(A0[k + 1]);
c00 = svmla_f32_x(body, c00, a0, b01);
c01 = svmla_f32_x(body, c01, a0, b11);
a0 = BroadcastBf16x2(A0 + k);
c00 = svbfdot_f32(c00, a0, b0);
c01 = svbfdot_f32(c01, a0, b1);
}
if (M > 1)
{
a0 = BroadcastBf16(A1[k + 0]);
c10 = svmla_f32_x(body, c10, a0, b00);
c11 = svmla_f32_x(body, c11, a0, b10);
a0 = BroadcastBf16(A1[k + 1]);
c10 = svmla_f32_x(body, c10, a0, b01);
c11 = svmla_f32_x(body, c11, a0, b11);
a0 = BroadcastBf16x2(A1 + k);
c10 = svbfdot_f32(c10, a0, b0);
c11 = svbfdot_f32(c11, a0, b1);
}
if (M > 2)
{
a0 = BroadcastBf16(A2[k + 0]);
c20 = svmla_f32_x(body, c20, a0, b00);
c21 = svmla_f32_x(body, c21, a0, b10);
a0 = BroadcastBf16(A2[k + 1]);
c20 = svmla_f32_x(body, c20, a0, b01);
c21 = svmla_f32_x(body, c21, a0, b11);
a0 = BroadcastBf16x2(A2 + k);
c20 = svbfdot_f32(c20, a0, b0);
c21 = svbfdot_f32(c21, a0, b1);
}
if (M > 3)
{
a0 = BroadcastBf16(A3[k + 0]);
c30 = svmla_f32_x(body, c30, a0, b00);
c31 = svmla_f32_x(body, c31, a0, b10);
a0 = BroadcastBf16(A3[k + 1]);
c30 = svmla_f32_x(body, c30, a0, b01);
c31 = svmla_f32_x(body, c31, a0, b11);
a0 = BroadcastBf16x2(A3 + k);
c30 = svbfdot_f32(c30, a0, b0);
c31 = svbfdot_f32(c31, a0, b1);
}
if (M > 4)
{
a0 = BroadcastBf16(A4[k + 0]);
c40 = svmla_f32_x(body, c40, a0, b00);
c41 = svmla_f32_x(body, c41, a0, b10);
a0 = BroadcastBf16(A4[k + 1]);
c40 = svmla_f32_x(body, c40, a0, b01);
c41 = svmla_f32_x(body, c41, a0, b11);
a0 = BroadcastBf16x2(A4 + k);
c40 = svbfdot_f32(c40, a0, b0);
c41 = svbfdot_f32(c41, a0, b1);
}
B0 += DF;
B1 += DF;
Expand Down Expand Up @@ -482,43 +468,31 @@ namespace Simd
}
for (size_t k = 0; k < K; k += 2)
{
svuint32_t b0u = svld1_u32(body, (const uint32_t*)B0);
b00 = svreinterpret_f32_u32(svlsl_n_u32_x(body, b0u, Base::Bf16::SHIFT));
b01 = svreinterpret_f32_u32(svand_u32_x(body, b0u, maskBits));
b0 = LoadBf16x2(B0, body);
if (M > 0)
{
a0 = BroadcastBf16(A0[k + 0]);
c00 = svmla_f32_x(body, c00, a0, b00);
a0 = BroadcastBf16(A0[k + 1]);
c00 = svmla_f32_x(body, c00, a0, b01);
a0 = BroadcastBf16x2(A0 + k);
c00 = svbfdot_f32(c00, a0, b0);
}
if (M > 1)
{
a0 = BroadcastBf16(A1[k + 0]);
c10 = svmla_f32_x(body, c10, a0, b00);
a0 = BroadcastBf16(A1[k + 1]);
c10 = svmla_f32_x(body, c10, a0, b01);
a0 = BroadcastBf16x2(A1 + k);
c10 = svbfdot_f32(c10, a0, b0);
}
if (M > 2)
{
a0 = BroadcastBf16(A2[k + 0]);
c20 = svmla_f32_x(body, c20, a0, b00);
a0 = BroadcastBf16(A2[k + 1]);
c20 = svmla_f32_x(body, c20, a0, b01);
a0 = BroadcastBf16x2(A2 + k);
c20 = svbfdot_f32(c20, a0, b0);
}
if (M > 3)
{
a0 = BroadcastBf16(A3[k + 0]);
c30 = svmla_f32_x(body, c30, a0, b00);
a0 = BroadcastBf16(A3[k + 1]);
c30 = svmla_f32_x(body, c30, a0, b01);
a0 = BroadcastBf16x2(A3 + k);
c30 = svbfdot_f32(c30, a0, b0);
}
if (M > 4)
{
a0 = BroadcastBf16(A4[k + 0]);
c40 = svmla_f32_x(body, c40, a0, b00);
a0 = BroadcastBf16(A4[k + 1]);
c40 = svmla_f32_x(body, c40, a0, b01);
a0 = BroadcastBf16x2(A4 + k);
c40 = svbfdot_f32(c40, a0, b0);
}
B0 += DF;
}
Expand Down