XeGPU 4kx4kx4k GEMM performance test case with prefetching enabaled.

intel · Dec 6, 2023 · efa1f09 · efa1f09
1 parent c05b357
commit efa1f09
Show file tree

Hide file tree

Showing 6 changed files with 523 additions and 12 deletions.
diff --git a/include/imex/ExecutionEngine/ImexRunnerUtils.h b/include/imex/ExecutionEngine/ImexRunnerUtils.h
@@ -51,8 +51,9 @@ extern "C" IMEX_RUNNERUTILS_EXPORT void
 _mlir_ciface_fillResource1DF32(MemRefDescriptor<float, 1> *ptr, // NOLINT
                                float value);
 extern "C" IMEX_RUNNERUTILS_EXPORT void
-_mlir_ciface_fillMatrixRandomBF16(MemRefDescriptor<bf16, 1> *ptr, // NOLINT
-                                  int nrows, int ncols);
+_mlir_ciface_fillMatrixRandomBF16(MemRefDescriptor<bf16, 1> *ptr);
+extern "C" IMEX_RUNNERUTILS_EXPORT void
+_mlir_ciface_fillMatrixRandomF16(MemRefDescriptor<f16, 1> *ptr);
 
 extern "C" IMEX_RUNNERUTILS_EXPORT void
 _mlir_ciface_printMemrefBF16(UnrankedMemRefType<bf16> *m);

diff --git a/lib/ExecutionEngine/ImexRunnerUtils.cpp b/lib/ExecutionEngine/ImexRunnerUtils.cpp
@@ -38,17 +38,28 @@ _mlir_ciface_fillResource1DF16(MemRefDescriptor<f16, 1> *ptr, // NOLINT
   std::fill_n(ptr->allocated, ptr->sizes[0], f16_val);
 }
 
-/// Fills the given 2D memref (i.e. matrix) passed as 1D memref
-// with randomly generated values. Numbers of rows and cols needs to be
-// specified and strides are assumed to be [ncols, 1]
+/// Fills 1D memref of bf16 type with random values uniformly
+/// distributed in the range (-0.5, 0.5)
 extern "C" void
-_mlir_ciface_fillMatrixRandomBF16(MemRefDescriptor<bf16, 1> *ptr, // NOLINT
-                                  int nrows, int ncols) {
+_mlir_ciface_fillMatrixRandomBF16(MemRefDescriptor<bf16, 1> *ptr) {
   std::random_device rd;
   std::mt19937 gen(rd());
   std::uniform_real_distribution<> dist(-0.5f, 0.5f);
 
-  for (int i = 0; i < nrows * ncols; i++) {
+  for (int i = 0; i < ptr->sizes[0]; i++) {
+    ptr->allocated[i] = dist(gen);
+  }
+}
+
+/// Fills 1D memref of f16 type with random values uniformly
+/// distributed in the range (-0.5, 0.5)
+extern "C" void
+_mlir_ciface_fillMatrixRandomF16(MemRefDescriptor<f16, 1> *ptr) {
+  std::random_device rd;
+  std::mt19937 gen(rd());
+  std::uniform_real_distribution<> dist(-0.5f, 0.5f);
+
+  for (int i = 0; i < ptr->sizes[0]; i++) {
     ptr->allocated[i] = dist(gen);
   }
 }