Browse Source

Dramatically simplify the dynamic intensity calculation by oversampling many runs through the opencl kernel till we're likely well within the timer resolution on windows.

nfactor-troky
Con Kolivas 12 years ago
parent
commit
7450b25e75
  1. 67
      driver-opencl.c
  2. 4
      miner.h

67
driver-opencl.c

@ -1463,9 +1463,7 @@ static void opencl_free_work(struct thr_info *thr, struct work *work) @@ -1463,9 +1463,7 @@ static void opencl_free_work(struct thr_info *thr, struct work *work)
const int thr_id = thr->id;
struct opencl_thread_data *thrdata = thr->cgpu_data;
_clState *clState = clStates[thr_id];
struct cgpu_info *gpu = thr->cgpu;
if (!gpu->dynamic)
clFinish(clState->commandQueue);
if (thrdata->res[FOUND]) {
@ -1496,7 +1494,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work, @@ -1496,7 +1494,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
_clState *clState = clStates[thr_id];
const cl_kernel *kernel = &clState->kernel;
const int dynamic_us = opt_dynamic_interval * 1000;
struct timeval tv_gpuend;
cl_int status;
size_t globalThreads[1];
@ -1504,9 +1501,26 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work, @@ -1504,9 +1501,26 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
int64_t hashes;
/* This finish flushes the readbuffer set with CL_FALSE later */
if (!gpu->dynamic)
clFinish(clState->commandQueue);
/* Windows' timer resolution is only 15ms so oversample 5x */
if (gpu->dynamic && (++gpu->intervals * dynamic_us) > 75) {
struct timeval tv_gpuend;
double gpu_us;
gettimeofday(&tv_gpuend, NULL);
gpu_us = us_tdiff(&tv_gpuend, &gpu->tv_gpustart) / gpu->intervals;
if (gpu_us > dynamic_us) {
if (gpu->intensity > MIN_INTENSITY)
--gpu->intensity;
} else if (gpu_us < dynamic_us / 2) {
if (gpu->intensity < MAX_INTENSITY)
++gpu->intensity;
}
memcpy(&(gpu->tv_gpustart), &tv_gpuend, sizeof(struct timeval));
gpu->intervals = 0;
}
set_threads_hashes(clState->vwidth, &hashes, globalThreads, localThreads[0], &gpu->intensity);
if (hashes > gpu->max_hashes)
gpu->max_hashes = hashes;
@ -1532,18 +1546,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work, @@ -1532,18 +1546,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
clFinish(clState->commandQueue);
}
if (gpu->dynamic) {
gettimeofday(&gpu->tv_gpumid, NULL);
if (gpu->new_work) {
gpu->new_work = false;
gpu->intervals = gpu->hit = 0;
}
if (!gpu->intervals) {
gpu->tv_gpustart.tv_sec = gpu->tv_gpumid.tv_sec;
gpu->tv_gpustart.tv_usec = gpu->tv_gpumid.tv_usec;
}
}
status = thrdata->queue_kernel_parameters(clState, &work->blk, globalThreads[0]);
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error: clSetKernelArg of all params failed.");
@ -1571,39 +1573,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work, @@ -1571,39 +1573,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
return -1;
}
if (gpu->dynamic) {
double gpu_us;
clFinish(clState->commandQueue);
/* Windows returns the same time for gettimeofday due to its
* 15ms timer resolution, so we must average the result over
* at least 5 values that are actually different to get an
* accurate result */
gpu->intervals++;
gettimeofday(&tv_gpuend, NULL);
gpu_us = us_tdiff(&tv_gpuend, &gpu->tv_gpumid);
if (gpu_us > 0 && ++gpu->hit > 4) {
gpu_us = us_tdiff(&tv_gpuend, &gpu->tv_gpustart) / gpu->intervals;
/* Very rarely we may get an overflow so put an upper
* limit on the detected time */
if (unlikely(gpu->gpu_us_average > 0 && gpu_us > gpu->gpu_us_average * 4))
gpu_us = gpu->gpu_us_average * 4;
gpu->gpu_us_average = (gpu->gpu_us_average + gpu_us * 0.63) / 1.63;
/* Try to not let the GPU be out for longer than
* opt_dynamic_interval in ms, but increase
* intensity when the system is idle in dynamic mode */
if (gpu->gpu_us_average > dynamic_us) {
if (gpu->intensity > MIN_INTENSITY)
--gpu->intensity;
} else if (gpu->gpu_us_average < dynamic_us / 2) {
if (gpu->intensity < MAX_INTENSITY)
++gpu->intensity;
}
gpu->intervals = gpu->hit = 0;
}
}
/* The amount of work scanned can fluctuate when intensity changes
* and since we do this one cycle behind, we increment the work more
* than enough to prevent repeating work */

4
miner.h

@ -401,9 +401,7 @@ struct cgpu_info { @@ -401,9 +401,7 @@ struct cgpu_info {
size_t shaders;
#endif
struct timeval tv_gpustart;
struct timeval tv_gpumid;
double gpu_us_average;
int intervals, hit;
int intervals;
#endif
bool new_work;

Loading…
Cancel
Save