Browse Source

nvml: store prev. clocks/limit for a proper reset on exit

master
Tanguy Pruvot 10 years ago
parent
commit
40fc2bdcbc
  1. 67
      nvml.cpp
  2. 4
      pools.cpp

67
nvml.cpp

@ -40,8 +40,9 @@ extern uint32_t device_mem_clocks[MAX_GPUS]; @@ -40,8 +40,9 @@ extern uint32_t device_mem_clocks[MAX_GPUS];
extern uint32_t device_plimit[MAX_GPUS];
extern int8_t device_pstate[MAX_GPUS];
uint8_t gpu_clocks_changed[MAX_GPUS] = { 0 };
uint8_t gpu_limits_changed[MAX_GPUS] = { 0 };
uint32_t clock_prev[MAX_GPUS] = { 0 };
uint32_t clock_prev_mem[MAX_GPUS] = { 0 };
uint32_t limit_prev[MAX_GPUS] = { 0 };
/*
* Wrappers to emulate dlopen() on other systems like Windows
@ -302,10 +303,6 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id) @@ -302,10 +303,6 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id)
if (n < 0 || n >= nvmlh->nvml_gpucount)
return -ENODEV;
// prevent double operations on the same gpu... to enhance
if (gpu_clocks_changed[dev_id])
return 0;
if (!device_gpu_clocks[dev_id] && !device_mem_clocks[dev_id])
return 0; // nothing to do
@ -314,6 +311,13 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id) @@ -314,6 +311,13 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id)
return -EPERM;
}
uint32_t mem_prev = clock_prev_mem[dev_id];
if (!mem_prev)
nvmlh->nvmlDeviceGetApplicationsClock(nvmlh->devs[n], NVML_CLOCK_MEM, &mem_prev);
uint32_t gpu_prev = clock_prev[dev_id];
if (!gpu_prev)
nvmlh->nvmlDeviceGetApplicationsClock(nvmlh->devs[n], NVML_CLOCK_GRAPHICS, &gpu_prev);
nvmlh->nvmlDeviceGetDefaultApplicationsClock(nvmlh->devs[n], NVML_CLOCK_MEM, &mem_clk);
rc = nvmlh->nvmlDeviceGetDefaultApplicationsClock(nvmlh->devs[n], NVML_CLOCK_GRAPHICS, &gpu_clk);
if (rc != NVML_SUCCESS) {
@ -332,7 +336,8 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id) @@ -332,7 +336,8 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id)
uint32_t nclocks = 0, clocks[127] = { 0 };
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, NULL);
nclocks = min(nclocks, 127);
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, clocks);
if (nclocks)
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, clocks);
for (int8_t u=0; u < nclocks; u++) {
// ordered by pstate (so highest is first memory clock - P0)
if (clocks[u] <= mem_clk) {
@ -344,7 +349,8 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id) @@ -344,7 +349,8 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id)
nclocks = 0;
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, NULL);
nclocks = min(nclocks, 127);
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, clocks);
if (nclocks)
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, clocks);
for (uint8_t u=0; u < nclocks; u++) {
// ordered desc, so get first
if (clocks[u] <= gpu_clk) {
@ -361,7 +367,9 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id) @@ -361,7 +367,9 @@ int nvml_set_clocks(nvml_handle *nvmlh, int dev_id)
return -1;
}
gpu_clocks_changed[dev_id] = 1;
// store previous clocks for reset on exit (or during wait...)
clock_prev[dev_id] = gpu_prev;
clock_prev_mem[dev_id] = mem_prev;
return 1;
}
@ -375,23 +383,24 @@ int nvml_reset_clocks(nvml_handle *nvmlh, int dev_id) @@ -375,23 +383,24 @@ int nvml_reset_clocks(nvml_handle *nvmlh, int dev_id)
if (n < 0 || n >= nvmlh->nvml_gpucount)
return -ENODEV;
if (gpu_clocks_changed[dev_id]) {
if (clock_prev[dev_id]) {
rc = nvmlh->nvmlDeviceResetApplicationsClocks(nvmlh->devs[n]);
if (rc != NVML_SUCCESS) {
applog(LOG_WARNING, "GPU #%d: unable to reset application clocks", dev_id);
}
gpu_clocks_changed[dev_id] = 0;
clock_prev[dev_id] = 0;
ret = 1;
}
if (gpu_limits_changed[dev_id]) {
uint32_t plimit = 0;
if (nvmlh->nvmlDeviceGetPowerManagementDefaultLimit) {
if (limit_prev[dev_id]) {
uint32_t plimit = limit_prev[dev_id];
if (nvmlh->nvmlDeviceGetPowerManagementDefaultLimit && !plimit) {
rc = nvmlh->nvmlDeviceGetPowerManagementDefaultLimit(nvmlh->devs[n], &plimit);
if (rc == NVML_SUCCESS)
nvmlh->nvmlDeviceSetPowerManagementLimit(nvmlh->devs[n], plimit);
} else if (plimit) {
rc = NVML_SUCCESS;
}
gpu_limits_changed[dev_id] = 0;
if (rc == NVML_SUCCESS)
nvmlh->nvmlDeviceSetPowerManagementLimit(nvmlh->devs[n], plimit);
ret = 1;
}
return ret;
@ -413,10 +422,6 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id) @@ -413,10 +422,6 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id)
if (device_pstate[dev_id] < 0)
return 0;
// prevent double operations on the same gpu... to enhance
if (gpu_clocks_changed[dev_id])
return 0;
if (nvmlh->app_clocks[n] != NVML_FEATURE_ENABLED) {
applog(LOG_WARNING, "GPU #%d: NVML app. clock feature is not allowed!", dev_id);
return -EPERM;
@ -438,9 +443,10 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id) @@ -438,9 +443,10 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id)
int8_t wanted_pstate = device_pstate[dev_id];
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, NULL);
nclocks = min(nclocks, 127);
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, clocks);
if (nclocks)
nvmlh->nvmlDeviceGetSupportedMemoryClocks(nvmlh->devs[n], &nclocks, clocks);
for (uint8_t u=0; u < nclocks; u++) {
// ordered by pstate (so high first)
// ordered by pstate (so highest P0 first)
if (u == wanted_pstate) {
mem_clk = clocks[u];
break;
@ -450,7 +456,8 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id) @@ -450,7 +456,8 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id)
nclocks = 0;
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, NULL);
nclocks = min(nclocks, 127);
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, clocks);
if (nclocks)
nvmlh->nvmlDeviceGetSupportedGraphicsClocks(nvmlh->devs[n], mem_clk, &nclocks, clocks);
for (uint8_t u=0; u < nclocks; u++) {
// ordered desc, so get first
if (clocks[u] <= gpu_clk) {
@ -468,7 +475,7 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id) @@ -468,7 +475,7 @@ int nvml_set_pstate(nvml_handle *nvmlh, int dev_id)
if (!opt_quiet)
applog(LOG_INFO, "GPU #%d: app clocks set to P%d (%u/%u)", dev_id, (int) wanted_pstate, mem_clk, gpu_clk);
gpu_clocks_changed[dev_id] = 1;
clock_prev[dev_id] = 1;
return 1;
}
@ -486,17 +493,17 @@ int nvml_set_plimit(nvml_handle *nvmlh, int dev_id) @@ -486,17 +493,17 @@ int nvml_set_plimit(nvml_handle *nvmlh, int dev_id)
if (!nvmlh->nvmlDeviceSetPowerManagementLimit)
return -ENOSYS;
uint32_t plimit = device_plimit[dev_id] * 1000U;
uint32_t pmin = 1000, pmax = 0;
uint32_t plimit = device_plimit[dev_id] * 1000;
uint32_t pmin = 1000, pmax = 0, prev_limit = 0;
if (nvmlh->nvmlDeviceGetPowerManagementLimitConstraints)
rc = nvmlh->nvmlDeviceGetPowerManagementLimitConstraints(nvmlh->devs[n], &pmin, &pmax);
if (rc != NVML_SUCCESS) {
if (!nvmlh->nvmlDeviceGetPowerManagementLimit)
return -ENOSYS;
pmax = 100 * 1000; // should not happen...
nvmlh->nvmlDeviceGetPowerManagementLimit(nvmlh->devs[n], &pmax);
}
nvmlh->nvmlDeviceGetPowerManagementLimit(nvmlh->devs[n], &prev_limit);
if (!pmax) pmax = prev_limit;
plimit = min(plimit, pmax);
plimit = max(plimit, pmin);
@ -511,7 +518,7 @@ int nvml_set_plimit(nvml_handle *nvmlh, int dev_id) @@ -511,7 +518,7 @@ int nvml_set_plimit(nvml_handle *nvmlh, int dev_id)
dev_id, plimit/1000U, pmin/1000U, pmax/1000U);
}
gpu_limits_changed[dev_id] = 1;
limit_prev[dev_id] = prev_limit;
return 1;
}

4
pools.cpp

@ -35,6 +35,7 @@ extern time_t firstwork_time; @@ -35,6 +35,7 @@ extern time_t firstwork_time;
extern volatile time_t g_work_time;
extern volatile int pool_switch_count;
extern volatile bool pool_is_switching;
extern uint8_t conditional_state[MAX_GPUS];
extern struct option options[];
@ -197,6 +198,9 @@ bool pool_switch(int pooln) @@ -197,6 +198,9 @@ bool pool_switch(int pooln)
// used to get the pool uptime
firstwork_time = time(NULL);
restart_threads();
// reset wait states
for (int n=0; n<opt_n_threads; n++)
conditional_state[n] = false;
// restore flags
allow_gbt = p->allow_gbt;

Loading…
Cancel
Save