Add sweep_utils.py script to tune heuristics (pytorch#3656)

Summary:

X-link: facebookresearch/FBGEMM#732

As title. heuristics tuning scripts for `bf16_fast_gemv`

currently the script needs a manual hack to update the kernel for passing in block dims to work. see comments in the code.

Reviewed By: ipiszy

Differential Revision: D68786295

fbgemm_gpu/experimental/gen_ai/src/quantize/fast_gemv/sweep_utils.py

@@ -0,0 +1,83 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch._inductor.utils import do_bench_using_profiling
+
+
+class SweepHeuristics:
+    def __init__(self):
+        self.m = 1
+        self.block_dims = [
+            (32, 1),
+            (32, 4),
+            (32, 8),
+            (32, 16),
+            (32, 32),
+            (64, 1),
+            (64, 2),
+            (64, 4),
+            (64, 8),
+            (64, 16),
+            (128, 1),
+            (128, 2),
+            (128, 4),
+            (128, 8),
+            (256, 1),
+            (256, 2),
+            (256, 4),
+            (512, 1),
+            (512, 2),
+            (1024, 1),
+        ]
+        self.nks = [(1280, 8192), (8192, 1024), (7168, 8192), (8192, 3584)]
+
+    def sweep_heuristics(self) -> None:
+        for n, k in self.nks:
+            x = torch.randn(size=(self.m, k), dtype=torch.half, device="cuda")
+            w = torch.randn(size=(n, k), dtype=torch.half, device="cuda")
+            best_elapsed_time, best_block_dim_x, best_block_dim_y = None, None, None
+
+            for block_dim_x, block_dim_y in self.block_dims:
+                if (
+                    (k % block_dim_x != 0)
+                    or (n % block_dim_x != 0)
+                    or ((k / block_dim_x) % 8 != 0)
+                ):
+                    continue
+                # This requires
+                # 1. update for testing purpose for `bf16_fast_gemv` pytorch custom op to pass in block_dim_x and block_dim_y
+                # 2. modify the fp16_fast_gemv.cu kernel signature to reflect the block_dim heuristics
+                # https://www.internalfb.com/code/fbsource/[bafd6390bc8c842b46d81be1a27dafd384503a53]/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py?lines=365
+                res = do_bench_using_profiling(
+                    lambda x=x, w=w, block_dim_x=block_dim_x, block_dim_y=block_dim_y: torch.ops.fbgemm.bf16_fast_gemv(
+                        x.T, w, block_dim_x, block_dim_y
+                    )
+                )
+
+                if best_elapsed_time is None or res < best_elapsed_time:
+                    best_elapsed_time, best_block_dim_x, best_block_dim_y = (
+                        res,
+                        block_dim_x,
+                        block_dim_y,
+                    )
+            if best_elapsed_time is None:
+                print("Error: No valid elapsed time found. Exiting the function.")
+                return
+            bw = (
+                (self.m * k * 2 + n * k * 2 + self.m * n * 2)
+                / (best_elapsed_time / 1000)
+                / (1024**3)
+            )
+
+            print(f"best elapsed time: {best_elapsed_time} ms")
+            print(f"best block_dim_x: {best_block_dim_x}")
+            print(f"best block_dim_y: {best_block_dim_y}")
+            print(f"best bw: {bw} GB/s")
+
+
+sweep_instance = SweepHeuristics()
+sweep_instance.sweep_heuristics()