Adding high bits for ATT buffer size (#171)

* Adding high bits for ATT buffer size

* Copilot review comments

* Add buffer limits

* Update src/pm4/sqtt_builder.h

---------

Co-authored-by: Giovanni <gbaraldi@amd.com>

[ROCm/aqlprofile commit: 0bb1a61e82]
This commit is contained in:
Baraldi, Giovanni
2025-07-23 00:53:55 +02:00
کامیت شده توسط GitHub
والد a75174f855
کامیت 4f8231fcc9
4فایلهای تغییر یافته به همراه29 افزوده شده و 32 حذف شده
@@ -222,6 +222,18 @@ class GpuSqttBuilder : public SqttBuilder, protected Primitives {
const uint64_t se_number_xcc = se_number_total / GetXCCNumber();
uint64_t base_addr = reinterpret_cast<uint64_t>(config->data_buffer_ptr);
const uint64_t base_step = GetBaseStep(config->data_buffer_size, config->se_mask);
// Old v1 API calls this with buffer == 0 first
if (config->data_buffer_size > 0)
{
// Max 16GB for gfx{9, 10, 12} and 512MB for gfx11. Min of 32 page per SE.
if (base_step >= (1ul<<34) || (Primitives::GFXIP_LEVEL == 11 && base_step >= (1ul<<29)))
throw std::runtime_error("SQTT Buffer size too high");
else if (base_step < (1ul<<17))
throw std::runtime_error("SQTT Buffer size too low");
}
config->capacity_per_se = base_step;
config->capacity_per_disabled_se = 1 << Primitives::TT_BUFF_ALIGN_SHIFT;
@@ -331,7 +343,7 @@ class GpuSqttBuilder : public SqttBuilder, protected Primitives {
const unsigned baddr_lo = Low32(base_addr >> Primitives::TT_BUFF_ALIGN_SHIFT);
const unsigned baddr_hi = High32(base_addr >> Primitives::TT_BUFF_ALIGN_SHIFT);
const uint32_t sqtt_size = bMaskedIn ? base_step : config->capacity_per_disabled_se;
const uint64_t sqtt_size = bMaskedIn ? base_step : config->capacity_per_disabled_se;
const uint32_t ctrl_val = Primitives::sqtt_ctrl_value(true);
Select_GRBM_SE_SH0(cmd_buffer, index);
@@ -432,7 +444,7 @@ class GpuSqttBuilder : public SqttBuilder, protected Primitives {
// Initialize cache flush request object
builder.BuildCacheFlushPacket(cmd_buffer, size_t(config->control_buffer_ptr),
config->control_buffer_size);
builder.BuildCacheFlushPacket(cmd_buffer, size_t(config->data_buffer_size),
builder.BuildCacheFlushPacket(cmd_buffer, size_t(config->data_buffer_ptr),
config->data_buffer_size);
// Program zero size of thread trace buffer
builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::SQ_THREAD_TRACE_SIZE_ADDR,
@@ -54,7 +54,7 @@ struct TraceConfig {
void* control_buffer_ptr = nullptr;
uint32_t control_buffer_size = 0;
void* data_buffer_ptr = nullptr;
uint32_t data_buffer_size = 0;
uint64_t data_buffer_size = 0;
// concurrent kernels mode
uint32_t concurrent = 0;