AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

PC Sampling: Implement lost sample count

Browse Source 
Change-Id: Idfdfbac71c1813dd7a97c301619cf8ce83713c53


[ROCm/ROCR-Runtime commit: 547c9cb143]
This commit is contained in:
David Yat Sin
2024-01-03 22:52:02 +00:00
parent 3d13db3c39
commit a84d407118
2 changed files with 24 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
+1
View File
@@ -734,6 +734,7 @@ class GpuAgent : public GpuAgentInt {
uint8_t* host_buffer_wrap_pos;
uint8_t* host_write_ptr;
uint8_t* host_read_ptr;
size_t lost_sample_count;
std::mutex host_buffer_mutex;
uint32_t which_buffer;
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+23 -7
View File
@@ -2606,6 +2606,7 @@ hsa_status_t GpuAgent::PcSamplingCreateFromId(HsaPcSamplingTraceId ioctlId,
return HSA_STATUS_ERROR;
}
ht_data.lost_sample_count = 0;
ht_data.host_buffer_wrap_pos = 0;
ht_data.host_write_ptr = ht_data.host_buffer;
ht_data.host_read_ptr = ht_data.host_write_ptr;
@@ -2913,8 +2914,10 @@ hsa_status_t GpuAgent::PcSamplingFlushHostTrapDeviceBuffers(
/* If the number of entries in old_val is larger than buf_size, then there was a buffer overflow
* and the 2nd level trap handler code will skip recording samples, causing lost samples
*/
if (*old_val > (uint64_t)ht_data.device_data->buf_size)
if (*old_val > (uint64_t)ht_data.device_data->buf_size) {
ht_data.lost_sample_count = *old_val - (uint64_t)ht_data.device_data->buf_size;
*old_val = (uint64_t)ht_data.device_data->buf_size;
}
to_copy = *old_val * session.sample_size();
@@ -3052,9 +3055,11 @@ void GpuAgent::PcSamplingThread() {
bytes_after_wrap = ht_data.host_write_ptr - host_buffer_begin;
while (bytes_before_wrap >= session.buffer_size()) {
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += session.buffer_size();
bytes_before_wrap = ht_data.host_buffer_wrap_pos - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
if (bytes_before_wrap + bytes_after_wrap >= session.buffer_size()) {
@@ -3064,13 +3069,16 @@ void GpuAgent::PcSamplingThread() {
bytes_before_wrap = 0;
ht_data.host_buffer_wrap_pos = 0;
bytes_after_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
while (bytes_after_wrap >= session.buffer_size()) {
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += session.buffer_size();
bytes_before_wrap = 0;
bytes_after_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
} else {
bytes_before_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
@@ -3078,9 +3086,11 @@ void GpuAgent::PcSamplingThread() {
while (bytes_before_wrap >= session.buffer_size()) {
assert(ht_data.host_read_ptr >= host_buffer_begin &&
ht_data.host_read_ptr + session.buffer_size() < host_buffer_end);
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, session.buffer_size(), NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += session.buffer_size();
bytes_before_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
}
}
@@ -3124,9 +3134,11 @@ hsa_status_t GpuAgent::PcSamplingFlush(pcs::PcsRuntime::PcSamplingSession& sessi
while (bytes_before_wrap > 0) {
size_t bytes_to_copy = std::min(bytes_before_wrap, session.buffer_size());
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += bytes_to_copy;
bytes_before_wrap = ht_data.host_buffer_wrap_pos - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
assert(ht_data.host_read_ptr == ht_data.host_buffer_wrap_pos);
@@ -3136,9 +3148,11 @@ hsa_status_t GpuAgent::PcSamplingFlush(pcs::PcsRuntime::PcSamplingSession& sessi
while (bytes_after_wrap > 0) {
size_t bytes_to_copy = std::min(bytes_after_wrap, session.buffer_size());
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += bytes_to_copy;
bytes_after_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
} else {
bytes_before_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
@@ -3148,9 +3162,11 @@ hsa_status_t GpuAgent::PcSamplingFlush(pcs::PcsRuntime::PcSamplingSession& sessi
assert(ht_data.host_read_ptr >= host_buffer_begin &&
ht_data.host_read_ptr + bytes_to_copy <= host_buffer_end);
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0, 0);
session.HandleSampleData(ht_data.host_read_ptr, bytes_to_copy, NULL, 0,
ht_data.lost_sample_count);
ht_data.host_read_ptr += bytes_to_copy;
bytes_before_wrap = ht_data.host_write_ptr - ht_data.host_read_ptr;
ht_data.lost_sample_count = 0;
}
}
return HSA_STATUS_SUCCESS;