wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

vx_spawn.c: Separate cluster-based scheduling code from original

Browse Source
This commit is contained in:
Hansung Kim
2024-03-26 16:36:57 -07:00
parent f050a08d77
commit 3729a05adc
2 changed files with 17 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
kernel/include/vx_spawn.h
View File

@@ -17,7 +17,9 @@
#include <stdint.h> #include <stdint.h>
#include <stdio.h> #include <stdio.h>
#ifndef CORES_PER_CLUSTER
#define CORES_PER_CLUSTER 2 #define CORES_PER_CLUSTER 2
#endif
#ifdef __cplusplus #ifdef __cplusplus
extern "C" { extern "C" {

23
kernel/src/vx_spawn.c
View File

@@ -128,7 +128,7 @@ static void __attribute__ ((noinline)) spawn_tasks_rem_stub() {
(p_wspawn_args->callback)(task_id, p_wspawn_args->arg); (p_wspawn_args->callback)(task_id, p_wspawn_args->arg);
} }
static void __attribute__ ((noinline)) spawn_tasks_all_cb() { static void __attribute__ ((noinline)) spawn_tasks_cluster_all_cb() {
// activate all threads // activate all threads
vx_tmc(-1); vx_tmc(-1);
@@ -140,6 +140,17 @@ static void __attribute__ ((noinline)) spawn_tasks_all_cb() {
vx_tmc_zero(); vx_tmc_zero();
} }
static void __attribute__ ((noinline)) spawn_tasks_all_cb() {
// activate all threads
vx_tmc(-1);
// call stub routine
spawn_tasks_all_stub();
// disable warp
vx_tmc_zero();
}
void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg) { void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg) {
// device specs // device specs
int NC = vx_num_cores(); int NC = vx_num_cores();
@@ -189,7 +200,7 @@ void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg
if (num_full_warps >= 1) { if (num_full_warps >= 1) {
// execute callback on other warps // execute callback on other warps
int nw = MIN(num_full_warps, NW); int nw = MIN(num_full_warps, NW);
vx_wspawn(nw, spawn_tasks_all_cb); vx_wspawn(nw, spawn_tasks_cluster_all_cb);
// activate all threads // activate all threads
vx_tmc(-1); vx_tmc(-1);
@@ -256,11 +267,7 @@ void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
rW = TW - fW * NW; // remaining warps rW = TW - fW * NW; // remaining warps
} }
int cluster_id = core_id / CORES_PER_CLUSTER; wspawn_tasks_args_t wspawn_args = { callback, arg, core_id * tasks_per_core, fW, rW };
const int tasks_per_cluster = tasks_per_core * CORES_PER_CLUSTER;
const int offset = cluster_id * tasks_per_cluster;
wspawn_tasks_args_t wspawn_args = { callback, arg, offset, fW, rW };
// wspawn_tasks_args_t wspawn_args = { callback, arg, core_id * tasks_per_core, fW, rW };
g_wspawn_args[core_id] = &wspawn_args; g_wspawn_args[core_id] = &wspawn_args;
if (TW >= 1) { if (TW >= 1) {
@@ -272,7 +279,7 @@ void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
vx_tmc(-1); vx_tmc(-1);
// call stub routine // call stub routine
spawn_tasks_cluster_all_stub(); spawn_tasks_all_stub();
// back to single-threaded // back to single-threaded
vx_tmc_one(); vx_tmc_one();