gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[PYTHON][TESTS][DOC] Various improvement of the API and code quality:

Browse Source 
* Simplified `triton.kernel` API to achieve lower latency:
  > .data_ptr() must now be passed as kernel argument. No more implicit
conversion from torch.tensor
  > compilation options are now constant attributes, i.e., opt.d('VAR')
becomes opt.VAR
  > torch.device must now be passed explicitly to triton.kernel (no
longer inferred from torch.tensor arguments)
* C++ tests moved to `python/tests/`
* C++ tutorial created in `tutorials/`
* Python tutorial created in python/tutorials/
* Version changed to 1.0alpha
* No longer copying C++ headers into the Python package
* added python/triton/ops/ package for pre-written Triton ops
This commit is contained in:
Philippe Tillet
2021-01-29 17:27:16 -05:00
parent a5a477c36b
commit 269ebc12e5
63 changed files with 2255 additions and 3883 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
python/triton/ops/matmul.c Normal file
View File

@@ -0,0 +1,97 @@
#define STM 8
#define STN 8
__global__ void matmul(TYPE * A __noalias __readonly __aligned(16),
TYPE * B __noalias __readonly __aligned(16),
TYPE * C __noalias __aligned(16),
float alpha,
int M,
int N,
int K __multipleof(16),
int lda __multipleof(LDA_POW2_DIV),
int ldb __multipleof(LDB_POW2_DIV),
int ldc __multipleof(LDC_POW2_DIV),
int* locks) {
// prologue
int pid = get_program_id(0);
int pidz = get_program_id(2);
int gridm = (M + TM - 1) / TM;
int gridn = (N + TN - 1) / TN;
// swizzle for better L2 performance
int width = STM*gridn;
int stm = pid / width;
int RSTM = min(gridm - stm*STM, STM);
int stn = (pid % width) / (RSTM*STN);
int RSTN = min(gridn - stn*STN, STN);
int laneid = pid % (RSTM * RSTN);
int lanem = laneid / RSTN;
int lanen = laneid % RSTN;
int pidm = stm*STM + lanem;
int pidn = stn*STN + lanen;
int rm[TM] = pidm * TM + 0 ... TM;
int rn[TN] = pidn * TN + 0 ... TN;
// split-k for better parrallelism
K = K / TZ;
int rk[TK] = 0 ... TK;
// pointers to operands
int offa[TM, TK] = (pidz*K + rk[newaxis, :]) * STRIDE_AK + rm[:, newaxis] * STRIDE_AM;
int offb[TK, TN] = (pidz*K + rk[:, newaxis]) * STRIDE_BK + rn[newaxis, :] * STRIDE_BN;
TYPE* pa[TM, TK] = A + offa;
TYPE* pb[TK, TN] = B + offb;
// prefetches operands
bool checka[TM, TK] = rk[newaxis, :] < K;
bool checkb[TK, TN] = rk[:, newaxis] < K;
TYPE a[TM, TK] = checka ? *pa : 0;
TYPE b[TK, TN] = checkb ? *pb : 0;
pa += TK * STRIDE_AK;
pb += TK * STRIDE_BK;
// reduction loop
float acc[TM, TN] = 0;
for(int k = K; k > 0; k -= TK){
#if (IS_TK_DIV_K==1)
bool checkk[TK] = k > TK;
#else
bool checkk[TK] = rk < k - TK;
#endif
bool checka[TM, TK] = checkk[newaxis, :];
bool checkb[TK, TN] = checkk[:, newaxis];
acc += a @ b;
#if (IS_TK_DIV_K==1)
a = *?(checka)pa;
b = *?(checkb)pb;
#else
a = checka ? *pa : 0;
b = checkb ? *pb : 0;
#endif
pa += TK * STRIDE_AK;
pb += TK * STRIDE_BK;
}
acc = acc * alpha;
TYPE c[TM, TN] = acc;
// epilogue
int rcm[TM] = pidm * TM + 0 ... TM;
int rcn[TN] = pidn * TN + 0 ... TN;
int offc[TM, TN] = rcm[:, newaxis] * ldc + rcn[newaxis, :];
TYPE* pc[TM, TN] = C + offc;
bool checkc[TM, TN] = rcm[:, newaxis] < M && rcn[newaxis, :] < N;
#if (TZ==1)
*?(checkc) pc = c;
#else
// accumulate partial result using spin-locks
int *plock = locks + rid;
int *pcount = plock + get_num_programs(0) * get_num_programs(1);
for(int repeat = 1; repeat == 1; repeat = atomic_cas(plock, 0, 1));
int count = *pcount;
if(count == 0)
*?(checkc) pc = c;
else
*?(checkc) pc = c + *?(checkc)pc;
atomic_xchg(pcount, (count + 1) % TZ);
atomic_xchg(plock, 0);
#endif
}