#include <stdlib.h>

Include dependency graph for elem_block_tune.h:

This graph shows which files directly or indirectly include this file:

Macros
#define	NEKO_EB_SWEEP_DEFAULT 0

#define	NEKO_TUNE_INIT 1.0e30f

#define	NEKO_TUNE_WARMUP 20

#define	NEKO_TUNE_TIME(T, LAUNCH, LX, C, ITERS)

#define	NEKO_TUNE_BEST(T, BEST, N)

#define	NEKO_TUNE_LOG(LX, T1, T2)

Functions
static int	neko_eb_sweep ()

static int	neko_eb_env ()

static int	neko_chunks_env ()

static int	neko_tune_rounds ()

static int	neko_tune_iters ()

Macro Definition Documentation

◆ NEKO_EB_SWEEP_DEFAULT

#define NEKO_EB_SWEEP_DEFAULT 0

Shared autotuner scaffolding for the SEM operator kernels

Each operator's tune() picks between the 1d variant and the kstep variant, and – when sweeping is enabled – between the elements per block candidates in elem_block.h. The sampling is interleaved and min reduced: timing the variants one after another in a fixed order lets clock drift bias the comparison by candidate position, which no amount of extra iterations removes.

A tune function using these macros is expected to have start, stop and stream in scope, plus a CASE_1D(LX) macro and a kstep launch macro taking (LX, C).

Definition at line 63 of file elem_block_tune.h.

◆ NEKO_TUNE_BEST

#define NEKO_TUNE_BEST	(	T,
		BEST,
		N
	)

Value:

  do {                                                                        \
    for (int c = 1; c < (N); c++) {                                           \
      if ((T)[c] < (T)[BEST]) { BEST = c; }                                   \
    }                                                                         \
  } while (0)

Definition at line 131 of file elem_block_tune.h.

◆ NEKO_TUNE_INIT

#define NEKO_TUNE_INIT 1.0e30f

Definition at line 66 of file elem_block_tune.h.

◆ NEKO_TUNE_LOG

#define NEKO_TUNE_LOG	(	LX,
		T1,
		T2
	)

Value:

  do {                                                                        \
    for (int c = 0; c < NEKO_CHUNKS_CANDIDATES; c++) {                        \
      if ((T1)[c] >= NEKO_TUNE_INIT) { continue; }                         \
      sprintf(neko_log_buf, "1D    ch=%-4d: %9.2f us/call",                   \
              NEKO_CHUNKS_SEL(LX, c), (T1)[c] * 10.0);                        \
      log_message(neko_log_buf);                                              \
    }                                                                         \
    for (int c = 0; c < NEKO_EB_CANDIDATES; c++) {                            \
      if ((T2)[c] >= NEKO_TUNE_INIT) { continue; }                         \
      sprintf(neko_log_buf, "KSTEP eb=%-4d: %9.2f us/call",                   \
              NEKO_EB_SEL(LX, c), (T2)[c] * 10.0);                            \
      log_message(neko_log_buf);                                              \
    }                                                                         \
  } while (0)

Definition at line 139 of file elem_block_tune.h.

◆ NEKO_TUNE_TIME

#define NEKO_TUNE_TIME	(	T,
		LAUNCH,
		LX,
		C,
		ITERS
	)

Value:

  do {                                                                        \
    float t_;                                                                 \
    HIP_CHECK(hipEventRecord(start, stream));                                           \
    for (int i_ = 0; i_ < (ITERS); i_++) { LAUNCH(LX, C); }                   \
    HIP_CHECK(hipEventRecord(stop, stream));                                            \
    HIP_CHECK(hipEventSynchronize(stop));                                               \
    HIP_CHECK(hipEventElapsedTime(&t_, start, stop));                                   \
    if (t_ < (T)[C]) { (T)[C] = t_; }                                         \
  } while (0)

Definition at line 120 of file elem_block_tune.h.

◆ NEKO_TUNE_WARMUP

#define NEKO_TUNE_WARMUP 20

Definition at line 67 of file elem_block_tune.h.

Function Documentation