Tianxing/moe gemm (#685)

Implemented moe gemm, test and benchmarking.

The gemm support weights so as the benchmark. The kernel uses the predefined block sizes.

The bench mark prints time, flops and memory throughput

Author: Chi-Chu319

.github/workflows/amd_perf_kernel_Integration_tests.yml

@@ -130,6 +130,7 @@ jobs:
           pytest -vvvv ./python/perf-kernels/softmax.py
           pytest -vvv ./python/perf-kernels/rmsnorm.py
           pytest -vvv ./python/perf-kernels/layernorm.py
+          pytest -vvv ./python/perf-kernels/fused_moe/moe-gemm.py
           sh ./python/perf-kernels/streamk/utils/unittest.sh
           pytest -vvv ./python/perf-kernels/multreduce_matmul_kernel.py
       - name: Run Perf Kernels Benchmark

python/perf-kernels/README.md

@@ -99,3 +99,6 @@ Kernel that implements RMS Norm over a row of tensor.
 
 ## `layernorm.py`
 Kernel that implements Layer Normalization over a row on tensor
+
+## `fused_moe/moe-gemm.py`
+Kernel that implements moe gemm.

python/perf-kernels/fused_moe/configs/device_name=AMD_Instinct_MI300X.json

@@ -0,0 +1,35 @@
+{
+  "small_M": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 4,
+    "num_warps": 8,
+    "num_stages": 2,
+    "waves_per_eu": 0,
+    "matrix_instr_nonkdim": 16,
+    "kpack": 2
+  },
+  "medium_M": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 2,
+    "waves_per_eu": 0,
+    "matrix_instr_nonkdim": 16,
+    "kpack": 2
+  },
+  "large_M": {
+    "BLOCK_SIZE_M": 256,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 2,
+    "waves_per_eu": 0,
+    "matrix_instr_nonkdim": 16,
+    "kpack": 2
+  }
+}