Gbaraldi/threadtrace2 (#724)

* Added first ATT API * Finalizing thread trace API * Fixing more rebase conflicts * Added codeobj disassembly sample * Fixing merge issues with rebase [2] * Adding ATT packets * Implemented thread trace intercept * Moved codeobj parser to same repo as rocprofiler * Moved thread trace to new API * Fixing merge conflicts * Fixing more merge conflicts * Adding thread trace packet reuse * Merged aql_profile_v2 headers * Linked ATT sample to aqlprofile * Updated decoder to include non-loaded codeobjs * Implemented ISA decoder into ATT sample * Added marker_id to vaddr * Updating aql_profile_v2 API to memcpy * Updating thread trace API to include 64bit markers. Using the result of ISA matching. * Added instruction type and cycles summary * Updated sample with selection of kernel by kernel_object * Added option to copy from memory kernels * Moved tool_data in thread_trace to dynamic alloc * Restoring hsa.cpp * Fixed ATT sample crash. General improvements. * Moved codeobj library to outside src/ * Updated license header * Moved codeobj_capture to camelcase * Solving some more merge conflicts * Update samples/advanced_thread_trace/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update samples/advanced_thread_trace/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update samples/code_object_isa_decode/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update source/lib/rocprofiler-sdk/thread_trace/CMakeLists.txt * Removing unused parameter check * Adding const to isEmpty * Removing unused warning * Adding libdw-dev to requirements * Running clang-format * Commenting out new aql calls * Clang format * Unused variable fix * Adding codeobj-decoder coverage * Commenting out threadtrace * Update samples/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * P * WOverloaded * Addressing clang-tidy * Virtual destructor on ttracer class * Corr id * Fixing code source format * Update CMakeLists.txt * Build fixes * Update source/lib/rocprofiler-sdk-codeobj/code_object_track.cpp Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Fix shadowing * Update CMakeLists.txt * Update samples/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Ammar ELWazir <ammar.elwazir@amd.com> Co-authored-by: Ammar ELWazir <aelwazir@amd.com> Co-authored-by: Benjamin Welton <bewelton@amd.com> [ROCm/rocprofiler-sdk commit: 69b8a43dc6]
2024-04-08 16:43:02 -03:00
@@ -0,0 +1,84 @@
+#pragma once
+
+#include "hip/hip_runtime.h"
+
+#define HIP_API_CALL(CALL)                                                                         \
+    {                                                                                              \
+        hipError_t error_ = (CALL);                                                                \
+        if(error_ != hipSuccess)                                                                   \
+        {                                                                                          \
+            lock_guard_t _hip_api_print_lk{print_lock};                                            \
+            fprintf(stderr,                                                                        \
+                    "%s:%d :: HIP error : %s\n",                                                   \
+                    __FILE__,                                                                      \
+                    __LINE__,                                                                      \
+                    hipGetErrorString(error_));                                                    \
+            exit(EXIT_FAILURE);                                                                    \
+        }                                                                                          \
+    }
+
+#define TILE_DIM 64
+
+template <typename T>
+__global__ void
+transposeNaive(T* odata, const T* idata, size_t size)
+{
+    size_t idx        = blockIdx.x * TILE_DIM + threadIdx.x;
+    size_t block_posy = blockIdx.y * TILE_DIM;
+
+    for(size_t idy = threadIdx.y; idy < TILE_DIM; idy += blockDim.y)
+        odata[size * idx + block_posy + idy] = idata[idx + (block_posy + idy) * size];
+}
+
+template <typename T>
+__global__ void
+transposeLdsNoBankConflicts(T* odata, const T* idata, size_t size)
+{
+    __shared__ T tile[TILE_DIM][TILE_DIM + 1];
+
+    size_t idx_in   = blockIdx.x * TILE_DIM + threadIdx.x;
+    size_t idy_in   = blockIdx.y * TILE_DIM + threadIdx.y;
+    size_t index_in = idx_in + idy_in * size;
+
+    size_t idx_out   = blockIdx.y * TILE_DIM + threadIdx.x;
+    size_t idy_out   = blockIdx.x * TILE_DIM + threadIdx.y;
+    size_t index_out = idx_out + idy_out * size;
+
+    for(size_t y = 0; y < TILE_DIM; y += blockDim.y)
+        tile[threadIdx.y + y][threadIdx.x] = idata[index_in + y * size];
+
+    __syncthreads();
+
+    for(size_t y = 0; y < TILE_DIM; y += blockDim.y)
+        odata[index_out + y * size] = tile[threadIdx.x][threadIdx.y + y];
+}
+
+// Generates more interesting ISA
+template <typename T>
+__global__ void
+transposeLdsSwapInplace(T* odata, const T* idata, size_t size)
+{
+    __shared__ T tile[TILE_DIM][TILE_DIM];
+
+    const size_t idx_in = blockIdx.x * TILE_DIM + threadIdx.x;
+
+    for(size_t idy = threadIdx.y; idy < TILE_DIM; idy += blockDim.y)
+        tile[idy][threadIdx.x] = idata[idx_in + (idy + blockIdx.y * TILE_DIM) * size];
+
+    __syncthreads();
+
+    for(size_t idy = threadIdx.y; idy < TILE_DIM; idy += blockDim.y)
+        if(idy < threadIdx.x)
+        {
+            T temp                 = tile[idy][threadIdx.x];
+            tile[idy][threadIdx.x] = tile[threadIdx.x][idy];
+            tile[threadIdx.x][idy] = temp;
+        }
+
+    __syncthreads();
+
+    const size_t idx_out = blockIdx.y * TILE_DIM + threadIdx.x;
+
+    for(size_t idy = threadIdx.y; idy < TILE_DIM; idy += blockDim.y)
+        odata[(blockIdx.x * TILE_DIM + idy) * size + idx_out] = tile[idy][threadIdx.x];
+}