Skip to content

TileOP: extend per-dimension B.DIM sweep to all four-branch interfaces #83

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
VV0003 merged 1 commit into from
Sep 7, 2026
Merged

TileOP: extend per-dimension B.DIM sweep to all four-branch interfaces #83

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
166 changes: 166 additions & 0 deletions test/tileop_api/src/ValidShapePerDimSweep.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,4 +123,170 @@ template <int VR,int VC> using T = Tile<Location::Vec, float, 32, 32, BLayout::R



namespace sw_a {
template <int VR,int VC> using T =
Tile<Location::Vec, float, 32, 32, BLayout::RowMajor, VR, VC>;
template <int VR,int VC> using TQ =
Tile<Location::Vec, int8_t, 32, 32, BLayout::RowMajor, VR, VC>;

// 三参同型: TPARTADD/MAX/MIN/MUL, TSEL(dst,mask,true), TFMA 4参
#define SW3(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &a,T<16,32> &b,T<16,32> &c){FN(a,b,c);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &a,T<16,-1> &b,T<16,-1> &c){FN(a,b,c);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &a,T<-1,32> &b,T<-1,32> &c){FN(a,b,c);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &a,T<-1,-1> &b,T<-1,-1> &c){FN(a,b,c);}
SW3(TPARTADD) SW3(TPARTMAX) SW3(TPARTMIN) SW3(TPARTMUL) SW3(TSEL)
#define SW4(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &a,T<16,32> &b,T<16,32> &c,T<16,32> &d){FN(a,b,c,d);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &a,T<16,-1> &b,T<16,-1> &c,T<16,-1> &d){FN(a,b,c,d);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &a,T<-1,32> &b,T<-1,32> &c,T<-1,32> &d){FN(a,b,c,d);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &a,T<-1,-1> &b,T<-1,-1> &c,T<-1,-1> &d){FN(a,b,c,d);}
SW4(TFMA)

// 二参 dst/src: TFILLPAD, TEXTRACT/TINSERT(带index)
#define SW2(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &a,T<16,32> &b){FN(a,b);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &a,T<16,-1> &b){FN(a,b);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &a,T<-1,32> &b){FN(a,b);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &a,T<-1,-1> &b){FN(a,b);}
SW2(TFILLPAD)
#define SW2I(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &a,T<16,32> &b){FN(a,b,0,0);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &a,T<16,-1> &b){FN(a,b,0,0);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &a,T<-1,32> &b){FN(a,b,0,0);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &a,T<-1,-1> &b){FN(a,b,0,0);}
SW2I(TEXTRACT) SW2I(TINSERT)
// TEXPANDS(dst, s)
#define SW1S(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &a){FN(a,1.0f);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &a){FN(a,1.0f);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &a){FN(a,1.0f);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &a){FN(a,1.0f);}
SW1S(TEXPANDS)
// TDEQUANT(dst,src,mult,zp)
__attribute__((noinline)) void tdq_ss(T<16,32> &d,TQ<16,32> &s){TDEQUANT(d,s,1.0f,2);}
__attribute__((noinline)) void tdq_sd(T<16,-1> &d,TQ<16,-1> &s){TDEQUANT(d,s,1.0f,2);}
__attribute__((noinline)) void tdq_ds(T<-1,32> &d,TQ<-1,32> &s){TDEQUANT(d,s,1.0f,2);}
__attribute__((noinline)) void tdq_dd(T<-1,-1> &d,TQ<-1,-1> &s){TDEQUANT(d,s,1.0f,2);}



} // namespace sw_a

namespace sw_b {
template <int VR,int VC> using T =
Tile<Location::Vec, float, 32, 32, BLayout::RowMajor, VR, VC>;
template <int VR,int VC> using C1 =
Tile<Location::Vec, float, 32, 1, BLayout::RowMajor, VR, VC>;

// EXPAND (dst=T, src=C1 单列)
#define SWEX(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &d,C1<16,1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &d,C1<16,-1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &d,C1<-1,1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &d,C1<-1,-1> &s){FN(d,s);}
SWEX(TROWEXPAND)
#define SWEX3(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &d,C1<16,1> &a,C1<16,1> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &d,C1<16,-1> &a,C1<16,-1> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &d,C1<-1,1> &a,C1<-1,1> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &d,C1<-1,-1> &a,C1<-1,-1> &b){FN(d,a,b);}
SWEX3(TROWEXPANDADD) SWEX3(TROWEXPANDDIV) SWEX3(TROWEXPANDMUL) SWEX3(TROWEXPANDSUB)
SWEX3(TROWEXPANDMAX) SWEX3(TROWEXPANDMIN) SWEX3(TROWEXPANDEXPDIF)

// ARGMAX/ARGMIN (dst=C1, src=T)
#define SWAR(FN) \
__attribute__((noinline)) void FN##_ss(C1<16,1> &d,T<16,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_sd(C1<16,-1> &d,T<16,-1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_ds(C1<-1,1> &d,T<-1,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_dd(C1<-1,-1> &d,T<-1,-1> &s){FN(d,s);}
SWAR(TROWARGMAX) SWAR(TROWARGMIN)



} // namespace sw_b

namespace sw_c {
template <int VR,int VC> using T =
Tile<Location::Vec, float, 32, 32, BLayout::RowMajor, VR, VC>;
template <int VR,int VC> using R1 =
Tile<Location::Vec, float, 1, 32, BLayout::RowMajor, VR, VC>;

// col expand: src 单行 R1, dst T
#define SWEX(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &d,R1<1,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &d,R1<1,-1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &d,R1<-1,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &d,R1<-1,-1> &s){FN(d,s);}
SWEX(TCOLEXPAND)
#define SWEX3(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &d,R1<1,32> &a,R1<1,32> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &d,R1<1,-1> &a,R1<1,-1> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &d,R1<-1,32> &a,R1<-1,32> &b){FN(d,a,b);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &d,R1<-1,-1> &a,R1<-1,-1> &b){FN(d,a,b);}
SWEX3(TCOLEXPANDADD) SWEX3(TCOLEXPANDDIV) SWEX3(TCOLEXPANDMUL) SWEX3(TCOLEXPANDSUB)
SWEX3(TCOLEXPANDMAX) SWEX3(TCOLEXPANDMIN) SWEX3(TCOLEXPANDEXPDIF)
#define SWAR(FN) \
__attribute__((noinline)) void FN##_ss(R1<1,32> &d,T<16,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_sd(R1<1,-1> &d,T<16,-1> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_ds(R1<-1,32> &d,T<-1,32> &s){FN(d,s);} \
__attribute__((noinline)) void FN##_dd(R1<-1,-1> &d,T<-1,-1> &s){FN(d,s);}
SWAR(TCOLARGMAX) SWAR(TCOLARGMIN)



} // namespace sw_c

namespace sw_d {
template <int VR,int VC> using T =
Tile<Location::Vec, float, 32, 32, BLayout::RowMajor, VR, VC>;
template <int VR,int VC> using O =
Tile<Location::Vec, int32_t, 32, 32, BLayout::RowMajor, VR, VC>;

// (dst, src, off) — off 的 ValidCol/Row 驱动四分支
#define SWGO(FN) \
__attribute__((noinline)) void FN##_ss(T<16,32> &d,T<16,32> &s,O<16,32> &o){FN(d,s,o);} \
__attribute__((noinline)) void FN##_sd(T<16,-1> &d,T<16,-1> &s,O<16,-1> &o){FN(d,s,o);} \
__attribute__((noinline)) void FN##_ds(T<-1,32> &d,T<-1,32> &s,O<-1,32> &o){FN(d,s,o);} \
__attribute__((noinline)) void FN##_dd(T<-1,-1> &d,T<-1,-1> &s,O<-1,-1> &o){FN(d,s,o);}
SWGO(TGATHER) SWGO(TSCATTER)



} // namespace sw_d

namespace sw_e {
using gms = global_tensor<float, RowMajor<32, 32>>;
template <int VR,int VC> using ST =
Tile<Location::Vec, float, 16, 32, BLayout::RowMajor, VR, VC>;

// Shared produced inline via TMOV_L2S_INSERT (never crosses the C++ ABI)
__attribute__((noinline)) void tst_ss(float *out, float *in) {
gms g(out); gms src(in); ST<16,32> t;
TLOAD(t, src);
auto sh = TMOV_L2S_INSERT(t);
TSTORE(g, sh);
}
__attribute__((noinline)) void tst_sd(float *out, float *in) {
gms g(out); gms src(in); ST<16,-1> t;
TLOAD(t, src);
auto sh = TMOV_L2S_INSERT(t);
TSTORE(g, sh);
}
__attribute__((noinline)) void tst_ds(float *out, float *in) {
gms g(out); gms src(in); ST<-1,32> t;
TLOAD(t, src);
auto sh = TMOV_L2S_INSERT(t);
TSTORE(g, sh);
}
__attribute__((noinline)) void tst_dd(float *out, float *in) {
gms g(out); gms src(in); ST<-1,-1> t;
TLOAD(t, src);
auto sh = TMOV_L2S_INSERT(t);
TSTORE(g, sh);
}


} // namespace sw_e

int main() { return 0; }
Loading