SWDEV-549717 - CP ROCm-7.0 (#366)

rocr: Increse SDMA queue size to 8MB

- When SDMA queue gets full when copying 2GB or more it blocks async
  copy api
- Improve/format logging

Co-authored-by: Saleel Kudchadker <Saleel.Kudchadker@amd.com>
Co-authored-by: Movsik, Valentin <Valentin.Movsik@amd.com>

Author: 3 people

runtime/hsa-runtime/core/runtime/amd_blit_sdma.cpp

@@ -71,7 +71,7 @@ inline uint32_t ptrhigh32(const void* p) {
 #endif
 }
 
-const size_t BlitSdmaBase::kQueueSize = 1024 * 1024;
+const size_t BlitSdmaBase::kQueueSize = 1024 * 1024 * 8;
 const size_t BlitSdmaBase::kCopyPacketSize = sizeof(SDMA_PKT_COPY_LINEAR);
 const size_t BlitSdmaBase::kMaxSingleCopySize = SDMA_PKT_COPY_LINEAR::kMaxSize_;
 const size_t BlitSdmaBase::kMaxSingleFillSize = SDMA_PKT_CONSTANT_FILL::kMaxSize_;
@@ -198,6 +198,8 @@ hsa_status_t BlitSdma<RingIndexTy, HwIndexMonotonic, SizeToCountOffset, useGCR>:
   if (agent_->driver().CreateQueue(agent_->node_id(), kQueueType_, 100, HSA_QUEUE_PRIORITY_MAXIMUM,
                                    rec_eng, queue_start_addr_, kQueueSize, nullptr,
                                    queue_resource_) != HSA_STATUS_SUCCESS) {
+    LogPrint(HSA_AMD_LOG_FLAG_INFO, "Failed to create queue, size=%d, type=%d,"
+       " priority=%d, engine_id=%d", kQueueSize, kQueueType_, HSA_QUEUE_PRIORITY_MAXIMUM, rec_eng);
     return HSA_STATUS_ERROR_OUT_OF_RESOURCES;
   }
 
@@ -568,7 +570,6 @@ hsa_status_t BlitSdma<RingIndexTy, HwIndexMonotonic, SizeToCountOffset,
   const size_t max_copy_size = max_single_linear_copy_size_ ? max_single_linear_copy_size_ :
                                kMaxSingleCopySize;
   const uint32_t num_copy_command = (size + max_copy_size - 1) / max_copy_size;
-
   // Assemble copy packets.
   std::vector<SDMA_PKT_COPY_LINEAR> buff(num_copy_command);
   BuildCopyCommand(reinterpret_cast<char*>(&buff[0]), num_copy_command, dst, src, size);

runtime/hsa-runtime/core/util/flag.cpp

@@ -49,22 +49,24 @@
 #include <string>
 #include <algorithm>
 #include <locale>
+#include <iomanip>
 
 namespace rocr {
 FILE* log_file = stderr;
 uint8_t log_flags[8];
 
 void log_printf(const char* file, int line, const char* format, ...) {
     va_list ap;
-    std::stringstream str_thrd_id;
-    str_thrd_id << std::hex << std::this_thread::get_id();
+    std::stringstream pidtid;
+    pidtid << "[pid:" << os::GetProcessId() << " tid: 0x" ;
+    pidtid << std::hex << std::setw(5) << std::this_thread::get_id() << "]";
     va_start(ap, format);
     char message[4096];
     vsnprintf(message, sizeof(message), format, ap);
     va_end(ap);
-    fprintf(log_file, ":%-25s:%-4d: %010lld us: [pid:%-5d tid:0x%s] [***rocr***] %s\n",
-            file, line, os::ReadAccurateClock()/1000ULL, os::GetProcessId(),
-            str_thrd_id.str().c_str(), message);
+    fprintf(log_file, ":7:%-25s:%-4d: %010lld us: %s [***rocr***] %s\n",
+            file, line, os::ReadAccurateClock()/1000ULL,
+            pidtid.str().c_str(), message);
     fflush(log_file);
 }
 

runtime/hsa-runtime/inc/hsa_ext_amd.h

@@ -3646,8 +3646,12 @@ hsa_status_t hsa_amd_queue_get_info(hsa_queue_t* queue, hsa_queue_info_attribute
  * @brief logging types
  */
 typedef enum hsa_amd_log_flag_s {
-   /* Log AQL packets internally enqueued by HSA for Blit Kernels */
+  /* Log AQL packets internally enqueued by HSA for Blit Kernels */
   HSA_AMD_LOG_FLAG_BLIT_KERNEL_PKTS = 0,
+  /* Log SDMA packets */
+  HSA_AMD_LOG_FLAG_BLIT_SDMA_PKTS = 1,
+  /* Log INFO */
+  HSA_AMD_LOG_FLAG_INFO = 2,
 } hsa_amd_log_flag_t;
 
 /**