Skip to content
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

ocl: updated tuned parameters, improved tuner, adjusted kernel #731

Merged
merged 1 commit into from
Nov 6, 2023
Merged

ocl: updated tuned parameters, improved tuner, adjusted kernel #731

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: 2 additions & 2 deletions src/acc/acc_bench_smm.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,8 +86,8 @@
#endif

#define ACC_BENCH_SMM_EPSILON(T) DBCSR_CONCATENATE(ACC_BENCH_SMM_EPSILON_, T)
#define ACC_BENCH_SMM_EPSILON_double 3E-3
#define ACC_BENCH_SMM_EPSILON_float 3E-3
#define ACC_BENCH_SMM_EPSILON_double 1E-3
#define ACC_BENCH_SMM_EPSILON_float 2E-3

#define ROUNDUP2(N, NPOT) ((((unsigned long long)N) + ((NPOT)-1)) & ~((NPOT)-1))
#define CHECK(EXPR, RPTR, VALUE) \
Expand Down
21 changes: 21 additions & 0 deletions src/acc/opencl/smm/README-bulktune.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,3 +66,24 @@ cp tune_multiply.csv smm/params/tune_multiply_P100.csv
```

Tuning kernels further is only sensible if the previously tuned parameters are embedded into the binary (such that the process does not start from scratch). Retuned parameters are captured with JSON-files as usual.

# Advanced Tuning

To utilize multiple devices per system and to accelerate tuning kernels, `tune_multiply.py` comes with built-in support for running under MPI (SPMD execution model). The basic assumption is to spawn one process per device usually with different kernels tuned per device (SPMD). Of course, tuning the same kernels in parallel on multiple devices is possible but it is a waste of resources. Tuning on multiple devices per system can be also more realistic given the common power budget of all devices and less room for an increased operating frequency per device (Turbo clock speed).

For example, a single dual-socket system with two PVC cards (modules) per socket exposes eight GPU devices (two GPU stacks or tiles per card). Then 350 kernels can be tuned in less than 2 1/2 hours with a duration of 200 seconds for tuning each kernel.

```bash
MAXTIME=200 NPARTS=8 UPDATE=1 JSONDIR=params/pvc mpirun \
./tune_multiply.sh -i 1 : \
./tune_multiply.sh -i 2 : \
./tune_multiply.sh -i 3 : \
./tune_multiply.sh -i 4 : \
./tune_multiply.sh -i 5 : \
./tune_multiply.sh -i 6 : \
./tune_multiply.sh -i 7 : \
./tune_multiply.sh -i 8 \
>out.log 2>&1
```

**NOTE**: The above shown example prefers environment variables over command-line options that would be common to the eight launches of `tune_multiply.sh`.
99 changes: 53 additions & 46 deletions src/acc/opencl/smm/kernels/multiply.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,13 +28,13 @@
#endif

#if !defined(AL) || (SM != SN) || (SM != BM) || (SN != SK) || (1 == BS)
# define ADX(M, K) adata[SM * K + M + a0]
# define BDX(K, N) bdata[SN * K + N + b0]
# define CDX(M, N) cdata[SM * N + M + c0]
# define ADX(M, K) adata[SM * K + M + a0] /* transposed */
# define BDX(K, N) bdata[SN * K + N + b0] /* linear */
# define CDX(M, N) cdata[SM * N + M + c0] /* transposed */
#else
# define ADX(M, K) adata[SK * M + K + a0]
# define BDX(K, N) bdata[SK * N + K + b0]
# define CDX(M, N) cdata[SN * M + N + c0]
# define ADX(M, K) adata[SK * M + K + a0] /* linear */
# define BDX(K, N) bdata[SK * N + K + b0] /* transposed */
# define CDX(M, N) cdata[SN * M + N + c0] /* linear */
#endif

#if defined(SLM_A)
Expand Down Expand Up @@ -69,6 +69,10 @@
# define CNM(N, M) cnm[M]
#endif

#if !defined(SINT) /* covers matrix shape */
# define SINT signed char
#endif

#if (1 == TN)
# define ZERO 0.f
#elif (3 == TN)
Expand Down Expand Up @@ -274,22 +278,22 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
int c0;
# if defined(SLM_C)
local T cnm[SN][SM + SLM_C - 1]; /* tile in SLM */
for (int n = idx; n < SN; n += SWG) {
for (SINT n = (SINT)idx; n < SN; n += SWG) {
UNROLL_FORCE(SM)
for (int m = 0; m < SM; ++m) cnm[n][m] = ZERO;
for (SINT m = 0; m < SM; ++m) cnm[n][m] = ZERO;
}
# elif (BM < SM || 1 != BN)
# if (1 != BN)
UNROLL(BN)
for (int bn = 0; bn < BN; ++bn)
for (SINT bn = 0; bn < BN; ++bn)
# endif
{
UNROLL_FORCE(BM)
for (int bm = 0; bm < BM; ++bm) CNM(bn, bm) = ZERO;
for (SINT bm = 0; bm < BM; ++bm) CNM(bn, bm) = ZERO;
}
# else
UNROLL_FORCE(SM)
for (int m = 0; m < SM; ++m) cnm[m] = ZERO;
for (SINT m = 0; m < SM; ++m) cnm[m] = ZERO;
# endif
# if defined(SLM_P)
UNROLL_FORCE(3 * BS)
Expand Down Expand Up @@ -332,7 +336,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
{
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) amk[m][k] = ADX(m, k);
for (SINT k = 0; k < SK; ++k) amk[m][k] = ADX(m, k);
}
}
#endif
Expand All @@ -345,7 +349,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
{
UNROLL(SK)
for (int k = 0; k < SK; ++k) bnk[n][k] = BDX(k, n);
for (SINT k = 0; k < SK; ++k) bnk[n][k] = BDX(k, n);
}
}
#elif defined(REG_B)
Expand All @@ -356,9 +360,9 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
{ /* copy or transpose B-matrix into registers */
# endif
UNROLL(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
# if (BM < SM || 1 != BN)
int bn = 0;
SINT bn = 0;
# if (1 != BN)
UNROLL_FORCE(BN)
for (; bn < BN; ++bn)
Expand Down Expand Up @@ -391,14 +395,14 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
UNROLL(BM)
# if (SM % BM)
for (int bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
# else
for (int bm = 0, m = m0; bm < BM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM; m = ++bm + m0)
# endif
{ /* general BK, A in registers */
int bn = 0;
SINT bn = 0;
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) amk[k] = ADX(m, k);
for (SINT k = 0; k < SK; ++k) amk[k] = ADX(m, k);
# if (1 != BN)
UNROLL(BN)
for (; bn < BN; ++bn)
Expand All @@ -419,7 +423,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
const int mc = bn, nc = idx;
# endif
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
CNM(nc, mc) = MAD(AMK(m, k),
# if defined(REG_B)
BNK(bn, k),
Expand Down Expand Up @@ -452,13 +456,13 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
T cnm[BM] = {ZERO}; /* column-block */
# endif
UNROLL(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
# if (SN % BN) || !defined(REG_B) || (defined(SLM_C) && (1 < BS)) || (1 == BS) || (1 != BN)
int bn = 0;
SINT bn = 0;
# endif
# if defined(REG_A) && !defined(SLM_A)
UNROLL_FORCE(BM)
for (int bm = 0; bm < BM; ++bm) amk[bm] = ADX(bm + m0, k);
for (SINT bm = 0; bm < BM; ++bm) amk[bm] = ADX(bm + m0, k);
# endif
# if (1 != BN)
UNROLL(BN)
Expand All @@ -479,9 +483,9 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
UNROLL_FORCE(BM)
# if (SM % BM)
for (int bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
# else
for (int bm = 0, m = m0; bm < BM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM; m = ++bm + m0)
# endif
{
# if defined(REG_A) && !defined(SLM_A)
Expand All @@ -497,7 +501,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
# if (1 == BS)
UNROLL(BM)
for (int bm = 0; bm < BM; ++bm) {
for (SINT bm = 0; bm < BM; ++bm) {
# if defined(ATOMIC_INC_NZ)
if (ZERO != CNM(idx, bm))
# endif
Expand All @@ -511,7 +515,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
}
# else /* general BK */
int bn = 0;
SINT bn = 0;
# if (1 != BN)
UNROLL(BN)
for (; bn < BN; ++bn)
Expand All @@ -529,9 +533,9 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
UNROLL(BM)
# if (SM % BM)
for (int bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
# else
for (int bm = 0, m = m0; bm < BM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM; m = ++bm + m0)
# endif
{
# if defined(SLM_C) && (1 < BS)
Expand All @@ -545,11 +549,11 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
const int nb = n;
# endif
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) CNM(nc, mc) = MAD(AMK(m, k), BNK(nb, k), CNM(nc, mc));
for (SINT k = 0; k < SK; ++k) CNM(nc, mc) = MAD(AMK(m, k), BNK(nb, k), CNM(nc, mc));
}
# if (1 == BS)
UNROLL(BM)
for (int bm = 0; bm < BM; ++bm) {
for (SINT bm = 0; bm < BM; ++bm) {
# if defined(ATOMIC_INC_NZ)
if (ZERO != CNM(idx, bm))
# endif
Expand All @@ -563,24 +567,24 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
# endif
}
#else
#else /* BM == SM && 1 == BN */
{ /* calculate result-tile using columns */
# if (1 == BS)
T cnm[UM] = {ZERO}; /* column-block */
# endif
# if (1 == BK)
UNROLL_OUTER(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
const T b = BNK(idx, k);
# if defined(SLM_A)
# if (WRK != SM)
for (int m = idx; m < SM; m += WRK) amk[m] = ADX(m, k);
for (SINT m = (SINT)idx; m < SM; m += WRK) amk[m] = ADX(m, k);
# else
amk[idx] = ADX(idx, k);
# endif
# elif defined(REG_A)
UNROLL_FORCE(SM)
for (int m = 0; m < SM; ++m) amk[m] = ADX(m, k);
for (SINT m = 0; m < SM; ++m) amk[m] = ADX(m, k);
# endif
# if defined(BARRIER) && (MAX(1, SGS) < SWG) && defined(SLM_A)
BARRIER(CLK_LOCAL_MEM_FENCE);
Expand All @@ -589,7 +593,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
const T a = AMK(idx, k);
# endif
UNROLL_FORCE(SM)
for (int m = 0; m < SM; ++m) {
for (SINT m = 0; m < SM; ++m) {
# if (200 /*CL_VERSION_2_0*/ <= __OPENCL_VERSION__) && !defined(SLM_A) && !defined(REG_A) && (WRK == SM) && \
(SM <= SGS || SM <= SWG)
# if (SM <= SGS)
Expand All @@ -609,7 +613,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
# if (1 == BS)
UNROLL(SM)
for (int m = 0; m < SM; ++m) {
for (SINT m = 0; m < SM; ++m) {
# if defined(ATOMIC_INC_NZ)
if (ZERO != CNM(idx, m))
# endif
Expand All @@ -620,7 +624,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
# endif
# else
int m = 0, u;
SINT m = 0, u;
# if (1 == UM)
UNROLL_OUTER(SM)
# endif
Expand All @@ -639,10 +643,10 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
# if defined(REG_A) && !defined(SLM_A)
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) amk[k] = ADX(um, k);
for (SINT k = 0; k < SK; ++k) amk[k] = ADX(um, k);
# endif
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
CNM(idx, vm) = MAD(AMK(um, k), BNK(idx, k), CNM(idx, vm));
}
}
Expand Down Expand Up @@ -677,10 +681,10 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
# if defined(REG_A) && !defined(SLM_A)
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) amk[k] = ADX(um, k);
for (SINT k = 0; k < SK; ++k) amk[k] = ADX(um, k);
# endif
UNROLL_FORCE(SK)
for (int k = 0; k < SK; ++k) {
for (SINT k = 0; k < SK; ++k) {
CNM(idx, vm) = MAD(AMK(um, k), BNK(idx, k), CNM(idx, vm));
}
}
Expand Down Expand Up @@ -709,7 +713,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
# endif
# if (BM < SM || 1 != BN)
{ /* atomically commit C-tile to global memory */
int bn = 0;
SINT bn = 0;
# if (1 != BN)
UNROLL(BN)
for (; bn < BN; ++bn)
Expand All @@ -722,9 +726,9 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
{
UNROLL_FORCE(BM)
# if (SM % BM)
for (int bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM && m < SM; m = ++bm + m0)
# else
for (int bm = 0, m = m0; bm < BM; m = ++bm + m0)
for (SINT bm = 0, m = m0; bm < BM; m = ++bm + m0)
# endif
{
# if defined(SLM_C)
Expand All @@ -744,7 +748,7 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
}
# else
{ /* atomically commit C-column to global memory */
int m = 0;
SINT m = 0;
# if defined(ATOMIC_ADD2_GLOBAL)
for (; m < (SM - 1); m += 2) {
# if defined(ATOMIC_INC_NZ)
Expand Down Expand Up @@ -774,6 +778,9 @@ FN(global T* restrict cdata, GLOBAL const T* restrict adata, GLOBAL const T* res
/* next iteration */
c0 = c1;
}
#endif
#if defined(BARRIER) && (MAX(1, SGS) < SWG) && defined(SLM_A) && (BM <= SM || 1 != BN || 1 != BK)
BARRIER(CLK_LOCAL_MEM_FENCE);
#endif
}
}
Loading
Loading