elem_block_tune.h File Reference

#include <stdlib.h>

Include dependency graph for elem_block_tune.h:

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

Go to the source code of this file.

Macros
#define	NEKO_EB_SWEEP_DEFAULT   1
#define	NEKO_TUNE_INIT   1.0e30f
#define	NEKO_TUNE_WARMUP   20
#define	NEKO_TUNE_FOR(C, ON, PIN, N)
#define	NEKO_TUNE_TIME(T, LAUNCH, LX, C, ITERS)
#define	NEKO_TUNE_US(T, ITERS)   ((T) * 1000.0 / (double) (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_pin ()
static int	neko_chunks_pin ()
static int	neko_tune_rounds ()
static int	neko_tune_iters ()

Macro Definition Documentation

NEKO_EB_SWEEP_DEFAULT

#define NEKO_EB_SWEEP_DEFAULT   1

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.

NEKO_AUTOTUNE takes a formulation out of the tuner's hands, but only the formulation: the geometry candidates of the one it names are still measured and reported, and it takes that formulation's own variable (NEKO_EB, NEKO_CHUNKS, ...) to fix the geometry too. Pinning the formulation used to imply candidate 0, which silently answered a different question than the one an A/B run of two formulations is asking. See NEKO_TUNE_FOR() below.

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

Definition at line 69 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 185 of file elem_block_tune.h.

NEKO_TUNE_FOR

#define NEKO_TUNE_FOR	(		C,
			ON,
			PIN,
			N
	)

Value:

  for (int C = (((PIN) >= 0) ? (PIN) : 0),                                    \
       C##_last_ = ((ON) ? (((PIN) >= 0) ? (PIN) + 1 : (N)) : 0);             \
       C < C##_last_; C++)

Definition at line 157 of file elem_block_tune.h.

NEKO_TUNE_INIT

#define NEKO_TUNE_INIT   1.0e30f

Definition at line 72 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), NEKO_TUNE_US((T1)[c], iters));          \
      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), NEKO_TUNE_US((T2)[c], iters));              \
      log_message(neko_log_buf);                                              \
    }                                                                         \
  } while (0)

Definition at line 193 of file elem_block_tune.h.

NEKO_TUNE_TIME

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

Value:

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

Definition at line 163 of file elem_block_tune.h.

NEKO_TUNE_US

#define NEKO_TUNE_US	(		T,
			ITERS
	)		((T) * 1000.0 / (double) (ITERS))

Definition at line 183 of file elem_block_tune.h.

NEKO_TUNE_WARMUP

#define NEKO_TUNE_WARMUP   20

Definition at line 73 of file elem_block_tune.h.

Function Documentation

neko_chunks_pin()

static int neko_chunks_pin ( )

static

Definition at line 112 of file elem_block_tune.h.

Here is the call graph for this function:

Here is the caller graph for this function:

neko_eb_pin()

static int neko_eb_pin ( )

static

Definition at line 95 of file elem_block_tune.h.

Here is the call graph for this function:

Here is the caller graph for this function:

neko_eb_sweep()

static int neko_eb_sweep ( )

static

Definition at line 75 of file elem_block_tune.h.

Here is the call graph for this function:

Here is the caller graph for this function:

neko_tune_iters()

static int neko_tune_iters ( )

static

Definition at line 136 of file elem_block_tune.h.

Here is the call graph for this function:

Here is the caller graph for this function:

neko_tune_rounds()

static int neko_tune_rounds ( )

static

Definition at line 128 of file elem_block_tune.h.

Here is the call graph for this function:

Here is the caller graph for this function: