This patch teaches the libgomp runtime how to probe the CUDA driver to
extract the number of Stream Multiprocessors that are available on the
graphics hardware and use that as the default value for num_gangs.
Without that patch, libgomp used to have num_gangs default to 32, which
was chosen arbitrarily. At least this value maps onto a hardware value.
More details regarding this patch can be found here:
https://gcc.gnu.org/ml/gcc-patches/2016-08/msg02064.html
https://gcc.gnu.org/ml/gcc-patches/2016-08/msg02084.html
Is this patch OK for trunk?
Cesar
2016-11-02 Cesar Philippidis <ce...@codesourcery.com>
Nathan Sidwell <nat...@acm.org>
gcc/
* config/nvptx/nvptx.c (PTX_GANG_DEFAULT): Set to zero.
libgomp/
* plugin/plugin-nvptx.c (nvptx_exec): Interrogate board attributes
to determine default geometry.
* testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c: Set gang
dimension.
diff --git a/gcc/config/nvptx/nvptx.c b/gcc/config/nvptx/nvptx.c
index 80fa9ae..782bbde 100644
--- a/gcc/config/nvptx/nvptx.c
+++ b/gcc/config/nvptx/nvptx.c
@@ -4174,7 +4174,7 @@ nvptx_expand_builtin (tree exp, rtx target, rtx ARG_UNUSED (subtarget),
/* Define dimension sizes for known hardware. */
#define PTX_VECTOR_LENGTH 32
#define PTX_WORKER_LENGTH 32
-#define PTX_GANG_DEFAULT 32
+#define PTX_GANG_DEFAULT 0 /* Defer to runtime. */
/* Validate compute dimensions of an OpenACC offload or routine, fill
in non-unity defaults. FN_LEVEL indicates the level at which a
diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c
index 327500c..91c1386 100644
--- a/libgomp/plugin/plugin-nvptx.c
+++ b/libgomp/plugin/plugin-nvptx.c
@@ -45,6 +45,7 @@
#include <stdio.h>
#include <unistd.h>
#include <assert.h>
+#include <errno.h>
static const char *
cuda_error (CUresult r)
@@ -932,9 +933,84 @@ nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
if (seen_zero)
{
+ /* See if the user provided GOMP_OPENACC_DIM environment
+ variable to specify runtime defaults. */
+ static int default_dims[GOMP_DIM_MAX];
+
+ if (!default_dims[0])
+ {
+ /* We only read the environment variable once. You can't
+ change it in the middle of execution. The sytntax is
+ the same as for the -fopenacc-dim compilation option. */
+ const char *env_var = getenv ("GOMP_OPENACC_DIM");
+ if (env_var)
+ {
+ const char *pos = env_var;
+
+ for (i = 0; *pos && i != GOMP_DIM_MAX; i++)
+ {
+ if (i && *pos++ != ':')
+ break;
+ if (*pos != ':')
+ {
+ const char *eptr;
+
+ errno = 0;
+ long val = strtol (pos, (char **)&eptr, 10);
+ if (errno || val < 0 || (unsigned)val != val)
+ break;
+ default_dims[i] = (int)val;
+ pos = eptr;
+ }
+ }
+ }
+
+ int warp_size, block_size, dev_size, cpu_size;
+ CUdevice dev = nvptx_thread()->ptx_dev->dev;
+ /* 32 is the default for known hardware. */
+ int gang = 0, worker = 32, vector = 32;
+
+ if (CUDA_SUCCESS == cuDeviceGetAttribute
+ (&block_size, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, dev)
+ && CUDA_SUCCESS == cuDeviceGetAttribute
+ (&warp_size, CU_DEVICE_ATTRIBUTE_WARP_SIZE, dev)
+ && CUDA_SUCCESS == cuDeviceGetAttribute
+ (&dev_size, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev)
+ && CUDA_SUCCESS == cuDeviceGetAttribute
+ (&cpu_size, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, dev))
+ {
+ GOMP_PLUGIN_debug (0, " warp_size=%d, block_size=%d,"
+ " dev_size=%d, cpu_size=%d\n",
+ warp_size, block_size, dev_size, cpu_size);
+ gang = (cpu_size / block_size) * dev_size;
+ worker = block_size / warp_size;
+ vector = warp_size;
+ }
+
+ /* There is no upper bound on the gang size. The best size
+ matches the hardware configuration. Logical gangs are
+ scheduled onto physical hardware. To maximize usage, we
+ should guess a large number. */
+ if (default_dims[GOMP_DIM_GANG] < 1)
+ default_dims[GOMP_DIM_GANG] = gang ? gang : 1024;
+ /* The worker size must not exceed the hardware. */
+ if (default_dims[GOMP_DIM_WORKER] < 1
+ || (default_dims[GOMP_DIM_WORKER] > worker && gang))
+ default_dims[GOMP_DIM_WORKER] = worker;
+ /* The vector size must exactly match the hardware. */
+ if (default_dims[GOMP_DIM_VECTOR] < 1
+ || (default_dims[GOMP_DIM_VECTOR] != vector && gang))
+ default_dims[GOMP_DIM_VECTOR] = vector;
+
+ GOMP_PLUGIN_debug (0, " default dimensions [%d,%d,%d]\n",
+ default_dims[GOMP_DIM_GANG],
+ default_dims[GOMP_DIM_WORKER],
+ default_dims[GOMP_DIM_VECTOR]);
+ }
+
for (i = 0; i != GOMP_DIM_MAX; i++)
- if (!dims[i])
- dims[i] = /* TODO */ 32;
+ if (!dims[i])
+ dims[i] = default_dims[i];
}
/* This reserves a chunk of a pre-allocated page of memory mapped on both
@@ -954,8 +1030,8 @@ nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
mapnum * sizeof (void *));
GOMP_PLUGIN_debug (0, " %s: kernel %s: launch"
" gangs=%u, workers=%u, vectors=%u\n",
- __FUNCTION__, targ_fn->launch->fn,
- dims[0], dims[1], dims[2]);
+ __FUNCTION__, targ_fn->launch->fn, dims[GOMP_DIM_GANG],
+ dims[GOMP_DIM_WORKER], dims[GOMP_DIM_VECTOR]);
// OpenACC CUDA
//
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c
index 8a755b8..3ca9388 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c
@@ -2,6 +2,8 @@
not optimized away at -O0, and then confuses the target assembler.
{ dg-skip-if "" { *-*-* } { "-O0" } { "" } } */
+/* { dg-additional-options "-fopenacc-dim=32" } */
+
#include <stdio.h>
#include <openacc.h>
