nvh::Profiler::SectionID ProfilerVK::beginSection(const char* name, VkCommandBuffer cmd)
{
nvh::Profiler::gpuTimeProvider_fn fnProvider = [&](SectionID i, uint32_t queryFrame, double& gpuTime) {
uint32_t idxBegin = getTimerIdx(i, queryFrame, true);
uint32_t idxEnd = getTimerIdx(i, queryFrame, false);
uint64_t times[2];
VkResult result = vkGetQueryPoolResults(m_device, m_queryPool, idxBegin, 2, sizeof(uint64_t) * 2, times, 0, VK_QUERY_RESULT_64_BIT);
if(result == VK_SUCCESS)
{
gpuTime = (double(times[1] - times[0]) * double(m_frequency)) / double(1000);
return true;
}
else
{
return false;
}
};
SectionID slot = Profiler::beginSection(name, "VK ", fnProvider);
if (getRequiredTimers() > m_queryPoolSize) {
resize();
}
uint32_t idx = getTimerIdx(slot, getSubFrame(), true);
// clear begin and end
vkCmdResetQueryPool(cmd, m_queryPool, idx, 2); // not ideal to do this per query
vkCmdWriteTimestamp(cmd, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, m_queryPool, idx);
return slot;
}
inline Profiler::Slot Profiler::beginSection( const char* name, GPUInterface* gpuif )
{
GLuint queryFrame = m_numFrames % FRAME_DELAY;
Slot slot = m_frameEntries++;
if (slot >= m_entries.size()){
grow((unsigned int)(m_entries.size() * 2));
if (gpuif){
gpuif->TimerGrow( getRequiredTimers() );
}
}
if (m_entries[slot].name != name ||
m_entries[slot].gpuif != gpuif )
{
m_entries[slot].name = name;
m_entries[slot].gpuif = gpuif;
m_resetDelay = CONFIG_DELAY;
}
int level = m_level++;
m_entries[slot].level = level;
m_entries[slot].splitter = false;
#ifdef SUPPORT_NVTOOLSEXT
{
nvtxEventAttributes_t eventAttrib = {0};
eventAttrib.version = NVTX_VERSION;
eventAttrib.size = NVTX_EVENT_ATTRIB_STRUCT_SIZE;
eventAttrib.colorType = NVTX_COLOR_ARGB;
unsigned char color[4];
color[0] = 255;
color[1] = 0;
color[2] = slot % 2 ? 127 : 255;
color[3] = 255;
color[2] -= level * 16;
color[3] -= level * 16;
eventAttrib.color = *(uint32_t*)(color);
eventAttrib.messageType = NVTX_MESSAGE_TYPE_ASCII;
eventAttrib.message.ascii = name;
nvtxRangePushEx(&eventAttrib);
}
#endif
if (gpuif){
gpuif->TimerSetup( getTimerIdx(slot,queryFrame,true) );
}
else{
glQueryCounter(m_entries[slot].queries[queryFrame],GL_TIMESTAMP);
}
m_entries[slot].deltas[queryFrame] = -getMicroSeconds();
return slot;
}
void ProfilerVK::resize()
{
if (getRequiredTimers() < m_queryPoolSize) return;
if (m_queryPool) {
// not exactly efficient, but when timers changed a lot, we have a slow frame anyway
// cleaner would be allocating more pools
vkDeviceWaitIdle(m_device);
vkDestroyQueryPool(m_device, m_queryPool, m_allocator);
}
VkQueryPoolCreateInfo create_info = { VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO };
create_info.queryType = VK_QUERY_TYPE_TIMESTAMP;
create_info.queryCount = getRequiredTimers();
m_queryPoolSize = create_info.queryCount;
VkResult res = vkCreateQueryPool(m_device, &create_info, m_allocator, &m_queryPool);
assert(res == VK_SUCCESS);
}
