void
pgstrom_perfmon_explain(pgstrom_perfmon *pfm, ExplainState *es)
{
double n = (double)pfm->num_samples;
char buf[256];
if (!pfm->enabled)
return;
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_to_load / 1000.0);
ExplainPropertyText("Total time to load", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_in_sendq / n / 1000.0);
ExplainPropertyText("Avg time in send-mq", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_kern_build / 1000.0);
ExplainPropertyText("Max time to build kernel", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_dma_send / n / 1000.0);
ExplainPropertyText("Avg time of DMA send", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_dma_send / 1000.0);
ExplainPropertyText("Total time of DMA send", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_kern_exec / n / 1000.0);
ExplainPropertyText("Avg time of kernel exec", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_kern_exec / 1000.0);
ExplainPropertyText("Total time of kernel exec", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_dma_recv / n / 1000.0);
ExplainPropertyText("Avg time of DMA recv", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_dma_recv / 1000.0);
ExplainPropertyText("Total time of DMA recv", buf, es);
snprintf(buf, sizeof(buf), "%.3f ms",
(double)pfm->time_in_recvq / n / 1000.0);
ExplainPropertyText("Avg time in recv-mq", buf, es);
}
/* ------------------------------------------------------------
*
* Routines copied from core PostgreSQL implementation
*
* ------------------------------------------------------------
*/
void
show_scan_qual(List *qual, const char *qlabel,
PlanState *planstate, List *ancestors,
ExplainState *es)
{
bool useprefix;
Node *node;
List *context;
char *exprstr;
useprefix = (IsA(planstate->plan, SubqueryScan) || es->verbose);
/* No work if empty qual */
if (qual == NIL)
return;
/* Convert AND list to explicit AND */
node = (Node *) make_ands_explicit(qual);
/* Set up deparsing context */
context = deparse_context_for_planstate((Node *) planstate,
ancestors,
es->rtable,
es->rtable_names);
/* Deparse the expression */
exprstr = deparse_expression(node, context, useprefix, false);
/* And add to es->str */
ExplainPropertyText(qlabel, exprstr, es);
}
void
show_device_kernel(Datum dprog_key, ExplainState *es)
{
StringInfoData str;
const char *kernel_source;
int32 extra_flags;
if (!es->verbose)
return;
kernel_source = pgstrom_get_devprog_kernel_source(dprog_key);
extra_flags = pgstrom_get_devprog_extra_flags(dprog_key);
initStringInfo(&str);
/*
* In case of EXPLAIN command context, we show the built-in logics
* like a usual #include preprocessor command.
* Practically, clCreateProgramWithSource() accepts multiple cstrings
* as if external files are included.
*/
if (extra_flags & DEVKERNEL_NEEDS_DEBUG)
appendStringInfo(&str, "#define PGSTROM_KERNEL_DEBUG 1\n");
appendStringInfo(&str, "#include \"opencl_common.h\"\n");
if (extra_flags & DEVFUNC_NEEDS_TIMELIB)
appendStringInfo(&str, "#include \"opencl_timelib.h\"\n");
if (extra_flags & DEVFUNC_NEEDS_TEXTLIB)
appendStringInfo(&str, "#include \"opencl_textlib.h\"\n");
if (extra_flags & DEVFUNC_NEEDS_NUMERICLIB)
appendStringInfo(&str, "#include \"opencl_numericlib.h\"\n");
if (extra_flags & DEVKERNEL_NEEDS_GPUSCAN)
appendStringInfo(&str, "#include \"opencl_gpuscan.h\"\n");
if (extra_flags & DEVKERNEL_NEEDS_GPUSORT)
appendStringInfo(&str, "#include \"opencl_gpusort.h\"\n");
if (extra_flags & DEVKERNEL_NEEDS_HASHJOIN)
appendStringInfo(&str, "#include \"opencl_hashjoin.h\"\n");
appendStringInfo(&str, "\n%s", kernel_source);
ExplainPropertyText("Kernel Source", str.data, es);
pfree(str.data);
}
void
pgstrom_perfmon_explain(pgstrom_perfmon *pfm, ExplainState *es)
{
bool multi_kernel = false;
char buf[256];
if (!pfm->enabled || pfm->num_samples == 0)
return;
/* common performance statistics */
ExplainPropertyInteger("number of requests", pfm->num_samples, es);
if (pfm->time_inner_load > 0)
{
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_inner_load));
ExplainPropertyText("total time for inner load", buf, es);
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_outer_load));
ExplainPropertyText("total time for outer load", buf, es);
}
else
{
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_outer_load));
ExplainPropertyText("total time to load", buf, es);
}
if (pfm->time_materialize > 0)
{
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_materialize));
ExplainPropertyText("total time to materialize", buf, es);
}
if (pfm->num_samples > 0 && (pfm->time_in_sendq > 0 ||
pfm->time_in_recvq > 0))
{
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_in_sendq /
(double)pfm->num_samples));
ExplainPropertyText("average time in send-mq", buf, es);
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_in_recvq /
(double)pfm->num_samples));
ExplainPropertyText("average time in recv-mq", buf, es);
}
if (pfm->time_kern_build > 0)
{
snprintf(buf, sizeof(buf), "%s",
usecond_unitary_format((double)pfm->time_kern_build));
ExplainPropertyText("max time to build kernel", buf, es);
}
if (pfm->num_dma_send > 0)
{
double band = (((double)pfm->bytes_dma_send * 1000000.0)
/ (double)pfm->time_dma_send);
snprintf(buf, sizeof(buf),
"%s/sec, len: %s, time: %s, count: %u",
bytesz_unitary_format(band),
bytesz_unitary_format((double)pfm->bytes_dma_send),
usecond_unitary_format((double)pfm->time_dma_send),
pfm->num_dma_send);
ExplainPropertyText("DMA send", buf, es);
}
if (pfm->num_dma_recv > 0)
{
double band = (((double)pfm->bytes_dma_recv * 1000000.0)
/ (double)pfm->time_dma_recv);
snprintf(buf, sizeof(buf),
"%s/sec, len: %s, time: %s, count: %u",
bytesz_unitary_format(band),
bytesz_unitary_format((double)pfm->bytes_dma_recv),
usecond_unitary_format((double)pfm->time_dma_recv),
pfm->num_dma_recv);
ExplainPropertyText("DMA recv", buf, es);
}
/* only gpupreagg */
if (pfm->num_kern_prep > 0)
{
multi_kernel = true;
snprintf(buf, sizeof(buf), "total: %s, avg: %s, count: %u",
usecond_unitary_format((double)pfm->time_kern_prep),
usecond_unitary_format((double)pfm->time_kern_prep /
(double)pfm->num_kern_prep),
pfm->num_kern_prep);
ExplainPropertyText("prep kernel exec", buf, es);
}
/* only gpupreagg */
if (pfm->num_kern_sort > 0)
{
multi_kernel = true;
snprintf(buf, sizeof(buf), "total: %s, avg: %s, count: %u",
usecond_unitary_format((double)pfm->time_kern_sort),
usecond_unitary_format((double)pfm->time_kern_sort /
(double)pfm->num_kern_sort),
pfm->num_kern_sort);
ExplainPropertyText("sort kernel exec", buf, es);
}
/* only gpuhashjoin */
if (pfm->num_kern_proj > 0)
{
multi_kernel = true;
snprintf(buf, sizeof(buf), "total: %s, avg: %s, count: %u",
usecond_unitary_format((double)pfm->time_kern_proj),
usecond_unitary_format((double)pfm->time_kern_proj /
(double)pfm->num_kern_proj),
pfm->num_kern_exec);
ExplainPropertyText("proj kernel exec", buf, es);
}
if (pfm->num_kern_exec > 0)
{
const char *label =
(multi_kernel ? "main kernel exec" : "kernel exec");
snprintf(buf, sizeof(buf), "total: %s, avg: %s, count: %u",
usecond_unitary_format((double)pfm->time_kern_exec),
usecond_unitary_format((double)pfm->time_kern_exec /
(double)pfm->num_kern_exec),
pfm->num_kern_exec);
ExplainPropertyText(label, buf, es);
}
/* for debugging if any */
if (pfm->time_debug1 > 0)
{
snprintf(buf, sizeof(buf), "debug1: %s",
usecond_unitary_format((double)pfm->time_debug1));
ExplainPropertyText("debug-1", buf, es);
}
if (pfm->time_debug2 > 0)
{
snprintf(buf, sizeof(buf), "debug2: %s",
usecond_unitary_format((double)pfm->time_debug2));
ExplainPropertyText("debug-2", buf, es);
}
if (pfm->time_debug3 > 0)
{
snprintf(buf, sizeof(buf), "debug3: %s",
usecond_unitary_format((double)pfm->time_debug3));
ExplainPropertyText("debug-3", buf, es);
}
if (pfm->time_debug4 > 0)
{
snprintf(buf, sizeof(buf), "debug4: %s",
usecond_unitary_format((double)pfm->time_debug4));
ExplainPropertyText("debug-4", buf, es);
}
}
/*
* twitterExplain
* Produce extra output for EXPLAIN
*/
static void
twitterExplain(ForeignScanState *node, ExplainState *es)
{
ExplainPropertyText("Twitter API", "Search", es);
}
