mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
26a91a9f04
Summary: Modified files in `benchmarks/tensorexpr` to add support for NVIDIA's Fuser for the jit compiler. This support has some modifications besides adding an option to support the NVIDIA fuser: * Adds FP16 Datatype support * Fixes SOL/Algo calculations to generally use the data type instead of being fixed to 4 bytes * Adds IR printing and kernel printing knobs * Adds a knob `input_iter` to create ranges of inputs currently only for reductions * Adds further reduction support for Inner and Outer dimension reductions that are compatible with the `input_iter` knob. * Added `simple_element`, `reduce2d_inner`, and `reduce2d_outer` to isolate performance on elementwise and reduction operations in the most minimal fashion. Pull Request resolved: https://github.com/pytorch/pytorch/pull/44101 Reviewed By: ngimel Differential Revision: D23713658 Pulled By: bertmaher fbshipit-source-id: d6b83cfab559aefe107c23b3c0f2df9923b3adc1
64 lines
1.6 KiB
Python
64 lines
1.6 KiB
Python
from . import benchmark
|
|
import numpy as np
|
|
|
|
|
|
class MatMulBench(benchmark.Benchmark):
|
|
def __init__(self, mode, device, dtype, B, M, N, K):
|
|
super().__init__(mode, device, dtype)
|
|
self.B = B
|
|
self.M = M
|
|
self.N = N
|
|
self.K = K
|
|
self.d1 = self.rand([B, M, N], device=device, dtype=dtype, requires_grad=self.requires_grad)
|
|
self.d2 = self.rand([B, N, K], device=device, dtype=dtype, requires_grad=self.requires_grad)
|
|
self.inputs = [self.d1, self.d2]
|
|
|
|
def forward(self, d1, d2):
|
|
y = self.matmul(d1, d2)
|
|
return y
|
|
|
|
def reference(self):
|
|
return np.matmul(self.numpy(self.d1), self.numpy(self.d2))
|
|
|
|
def config(self):
|
|
return [self.B, self.M, self.N, self.K]
|
|
|
|
@staticmethod
|
|
def module():
|
|
return "batch_matmul"
|
|
|
|
def memory_workload(self):
|
|
if self.mode == "fwd":
|
|
sol_count = 1
|
|
algorithmic_count = 1
|
|
else:
|
|
sol_count = 1 + 1
|
|
algorithmic_count = 1 + (1 + 1)
|
|
|
|
buffer_size = (
|
|
self.B * self.M * self.N
|
|
+ self.B * self.M * self.N
|
|
+ self.B * self.N * self.K
|
|
)
|
|
return {
|
|
"sol": buffer_size * sol_count,
|
|
"algorithmic": buffer_size * algorithmic_count,
|
|
}
|
|
|
|
def compute_workload(self):
|
|
if self.mode == "fwd":
|
|
count = 1
|
|
else:
|
|
count = 1 + (1 + 1)
|
|
|
|
op_count = 2 * self.B * self.M * self.N * self.K
|
|
|
|
return op_count * count
|
|
|
|
@staticmethod
|
|
def default_configs():
|
|
return [[128, 64, 128, 256]]
|
|
|
|
|
|
benchmark.register_benchmark_class(MatMulBench)
|