Remove triton.ops dependency from fbgemm (pytorch#3329)

Summary:
X-link: facebookresearch/FBGEMM#423

Pull Request resolved: pytorch#3329

3.2.x moves triton.ops to a seperate github repo. Since fbgemm only depends on triton.ops in a single place, we can simply copy-paste the upstream file.

Copied the file from here: https://github.com/triton-lang/kernels/blob/8821ef322394ee2d3c58a859780ee1e2e10b5c79/kernels/matmul_perf_model.py

Reviewed By: bertmaher, q10, davidberard98

Differential Revision: D65496960

fbshipit-source-id: f0a959bb98888964e441da292f494352485d3b15

fbgemm_gpu/experimental/gemm/CMakeLists.txt

@@ -10,7 +10,8 @@
 
 set(experimental_triton_python_source_files
     triton_gemm/__init__.py
-    triton_gemm/fp8_gemm.py)
+    triton_gemm/fp8_gemm.py
+    triton_gemm/matmul_perf_model.py)
 
 
 ################################################################################

fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py

@@ -13,13 +13,14 @@
 import triton  # @manual
 
 import triton.language as tl  # @manual
-from torch._tensor import Tensor
 
-from triton import Config  # @manual
-from triton.ops.matmul_perf_model import (  # @manual
+from fbgemm_gpu.experimental.gemm.triton_gemm.matmul_perf_model import (
     early_config_prune,
     estimate_matmul_time,
 )
+from torch._tensor import Tensor
+
+from triton import Config  # @manual
 from triton.runtime.jit import reinterpret as tl_reinterpret, TensorWrapper  # @manual
 
 logger: logging.Logger = logging.getLogger(__name__)

fbgemm_gpu/experimental/gemm/triton_gemm/matmul_perf_model.py

@@ -0,0 +1,232 @@
+# Source: https://github.com/triton-lang/kernels/blob/8821ef322394ee2d3c58a859780ee1e2e10b5c79/kernels/matmul_perf_model.py
+
+# This file is taken from the upstream triton-lang/kernels repo.
+# Currently that repo does not have a license file, so disabling
+# the license lint for now:
+# @lint-ignore-every LICENSELINT
+
+# flake8: noqa
+# pyre-ignore-all-errors
+
+import functools
+import heapq
+
+import torch
+
+from triton import cdiv  # @manual
+from triton.runtime import driver  # @manual
+from triton.testing import (  # @manual
+    get_dram_gbps,
+    get_max_simd_tflops,
+    get_max_tensorcore_tflops,
+    nvsmi,
+)
+
+
+@functools.lru_cache()
+def get_clock_rate_in_khz():
+    try:
+        return nvsmi(["clocks.max.sm"])[0] * 1e3
+    except FileNotFoundError:
+        import pynvml  # @manual=fbsource//third-party/pypi/pynvml:pynvml
+
+        pynvml.nvmlInit()
+        handle = pynvml.nvmlDeviceGetHandleByIndex(0)
+        return pynvml.nvmlDeviceGetMaxClockInfo(handle, pynvml.NVML_CLOCK_SM) * 1e3
+
+
+def get_tensorcore_tflops(device, num_ctas, num_warps, dtype):
+    """return compute throughput in TOPS"""
+    total_warps = num_ctas * min(num_warps, 4)
+    num_subcores = (
+        driver.active.utils.get_device_properties(device)["multiprocessor_count"] * 4
+    )  # on recent GPUs
+    tflops = (
+        min(num_subcores, total_warps)
+        / num_subcores
+        * get_max_tensorcore_tflops(dtype, get_clock_rate_in_khz(), device)
+    )
+    return tflops
+
+
+def get_simd_tflops(device, num_ctas, num_warps, dtype):
+    """return compute throughput in TOPS"""
+    total_warps = num_ctas * min(num_warps, 4)
+    num_subcores = (
+        driver.active.utils.get_device_properties(device)["multiprocessor_count"] * 4
+    )  # on recent GPUs
+    tflops = (
+        min(num_subcores, total_warps)
+        / num_subcores
+        * get_max_simd_tflops(dtype, get_clock_rate_in_khz(), device)
+    )
+    return tflops
+
+
+def get_tflops(device, num_ctas, num_warps, dtype):
+    capability = torch.cuda.get_device_capability(device)
+    if capability[0] < 8 and dtype == torch.float32:
+        return get_simd_tflops(device, num_ctas, num_warps, dtype)
+    return get_tensorcore_tflops(device, num_ctas, num_warps, dtype)
+
+
+def estimate_matmul_time(
+    # backend, device,
+    num_warps,
+    num_stages,  #
+    A,
+    B,
+    C,  #
+    M,
+    N,
+    K,  #
+    BLOCK_M,
+    BLOCK_N,
+    BLOCK_K,
+    SPLIT_K,  #
+    debug=False,
+    **kwargs,  #
+):
+    """return estimated running time in ms
+    = max(compute, loading) + store"""
+    device = torch.cuda.current_device()
+    dtype = A.dtype
+    dtsize = A.element_size()
+
+    num_cta_m = cdiv(M, BLOCK_M)
+    num_cta_n = cdiv(N, BLOCK_N)
+    num_cta_k = SPLIT_K
+    num_ctas = num_cta_m * num_cta_n * num_cta_k
+
+    # If the input is smaller than the block size
+    M, N = max(M, BLOCK_M), max(N, BLOCK_N)
+
+    # time to compute
+    total_ops = 2 * M * N * K / (1024 * 1024 * 1024)  # GOPS
+    tput = get_tflops(device, num_ctas, num_warps, dtype)
+    compute_ms = total_ops / tput
+
+    # time to load data
+    num_sm = driver.active.utils.get_device_properties(device)["multiprocessor_count"]
+    active_cta_ratio = min(1, num_ctas / num_sm)
+    active_cta_ratio_bw1 = min(
+        1, num_ctas / 32
+    )  # 32 active ctas are enough to saturate
+    active_cta_ratio_bw2 = max(
+        min(1, (num_ctas - 32) / (108 - 32)), 0
+    )  # 32-108, remaining 5%
+    dram_bw = get_dram_gbps(device) * (
+        active_cta_ratio_bw1 * 0.95 + active_cta_ratio_bw2 * 0.05
+    )  # in GB/s
+    l2_bw = dram_bw * 4  # rough estimation (should be 4.7 for A100?)
+    # assume 80% of (following) loads are in L2 cache
+    load_a_dram = M * K * dtsize * (1 + 0.2 * (num_cta_n - 1))
+    load_a_l2 = M * K * dtsize * 0.8 * (num_cta_n - 1)
+    load_b_dram = N * K * dtsize * (1 + 0.2 * (num_cta_m - 1))
+    load_b_l2 = N * K * dtsize * 0.8 * (num_cta_m - 1)
+    # total
+    total_dram = (load_a_dram + load_b_dram) / (1024 * 1024)  # MB
+    total_l2 = (load_a_l2 + load_b_l2) / (1024 * 1024)
+    # loading time in ms
+    load_ms = total_dram / dram_bw + total_l2 / l2_bw
+
+    # estimate storing time
+    store_bw = dram_bw * 0.6  # :o
+    store_c_dram = M * N * dtsize * SPLIT_K / (1024 * 1024)  # MB
+    if SPLIT_K == 1:
+        store_ms = store_c_dram / store_bw
+    else:
+        reduce_bw = store_bw
+        store_ms = store_c_dram / reduce_bw
+        # c.zero_()
+        zero_ms = M * N * 2 / (1024 * 1024) / store_bw
+        store_ms += zero_ms
+
+    total_time_ms = max(compute_ms, load_ms) + store_ms
+    if debug:
+        print(
+            f"Total time: {total_time_ms}ms, compute time: {compute_ms}ms, "
+            f"loading time: {load_ms}ms, store time: {store_ms}ms, "
+            f"Activate CTAs: {active_cta_ratio*100}%"
+        )
+    return total_time_ms
+
+
+def early_config_prune(configs, named_args, **kwargs):
+    device = torch.cuda.current_device()
+    capability = torch.cuda.get_device_capability()
+    # BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, num_warps, num_stages
+    dtsize = named_args["A"].element_size()
+    dtype = named_args["A"].dtype
+
+    # 1. make sure we have enough smem
+    pruned_configs = []
+    for config in configs:
+        kw = config.kwargs
+        BLOCK_M, BLOCK_N, BLOCK_K, num_stages = (
+            kw["BLOCK_M"],
+            kw["BLOCK_N"],
+            kw["BLOCK_K"],
+            config.num_stages,
+        )
+
+        max_shared_memory = driver.active.utils.get_device_properties(device)[
+            "max_shared_mem"
+        ]
+        required_shared_memory = (BLOCK_M + BLOCK_N) * BLOCK_K * num_stages * dtsize
+        if required_shared_memory <= max_shared_memory:
+            pruned_configs.append(config)
+    configs = pruned_configs
+
+    # Some dtypes do not allow atomic_add
+    if dtype not in [torch.float16, torch.float32]:
+        configs = [config for config in configs if config.kwargs["SPLIT_K"] == 1]
+
+    # group configs by (BLOCK_M,_N,_K, SPLIT_K, num_warps)
+    configs_map = {}
+    for config in configs:
+        kw = config.kwargs
+        BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, num_warps, num_stages = (
+            kw["BLOCK_M"],
+            kw["BLOCK_N"],
+            kw["BLOCK_K"],
+            kw["SPLIT_K"],
+            config.num_warps,
+            config.num_stages,
+        )
+
+        key = (BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, num_warps)
+        if key in configs_map:
+            configs_map[key].append((config, num_stages))
+        else:
+            configs_map[key] = [(config, num_stages)]
+
+    pruned_configs = []
+    for k, v in configs_map.items():
+        BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, num_warps = k
+        if capability[0] >= 8:
+            # compute cycles (only works for ampere GPUs)
+            mmas = BLOCK_M * BLOCK_N * BLOCK_K / (16 * 8 * 16)
+            mma_cycles = mmas / min(4, num_warps) * 8
+
+            ldgsts_latency = 300  # Does this matter?
+            optimal_num_stages = ldgsts_latency / mma_cycles
+
+            # nearest stages, prefer large #stages
+            nearest = heapq.nsmallest(
+                2,
+                v,
+                key=lambda x: (
+                    10 + abs(x[1] - optimal_num_stages)
+                    if (x[1] - optimal_num_stages) < 0
+                    else x[1] - optimal_num_stages
+                ),
+            )
+
+            for n in nearest:
+                pruned_configs.append(n[0])
+        else:  # Volta & Turing only supports num_stages <= 2
+            random_config = v[0][0]
+            random_config.num_stages = 2
+            pruned_configs.append(random_config)
+    return pruned_configs