pytorch

2c2e7f63 - unify compare kernels (#69111)

unify compare kernels (#69111) Summary: This unifies 6 compare ops (NE, EQ, LT, LE, GE, GT) into 2 kernels, reducing context size. Performance is ~5% worse for low width broadcasted cases, on-par for non-broadcasted With this PR, benchmarks for contiguous, 1M-MM, 1M-M1, op with scalar (size in MB and bandwidth in GB/s): ``` 5.0, 795.9 10.0, 650.5 15.0, 706.2 20.0, 731.6 25.0, 744.9 30.0, 758.1 35.0, 762.6 40.0, 768.8 45.0, 775.7 50.0, 780.7 55.0, 781.7 60.0, 783.0 65.0, 784.8 70.0, 790.7 75.0, 789.2 80.0, 794.4 85.0, 794.2 90.0, 797.4 95.0, 796.3 100.0, 798.0 3.0, 363.7 1.0, 122.2 3.0, 385.5 6.0, 420.4 2.0, 142.9 6.0, 755.5 9.0, 438.3 3.0, 151.6 9.0, 684.5 12.0, 449.5 4.0, 156.4 12.0, 702.9 15.0, 463.7 5.0, 159.6 15.0, 716.8 18.0, 472.7 6.0, 161.4 18.0, 737.0 21.0, 477.6 7.0, 162.4 21.0, 745.6 24.0, 480.9 8.0, 164.1 24.0, 755.4 27.0, 483.7 9.0, 163.7 27.0, 760.7 30.0, 487.3 10.0, 165.9 30.0, 770.4 33.0, 491.4 11.0, 166.3 33.0, 774.3 36.0, 492.9 12.0, 166.2 36.0, 779.0 39.0, 494.7 13.0, 166.7 39.0, 782.5 42.0, 491.3 14.0, 166.7 42.0, 789.0 45.0, 495.1 15.0, 167.5 45.0, 790.0 48.0, 499.7 16.0, 167.7 48.0, 791.8 51.0, 496.2 17.0, 166.9 51.0, 794.0 54.0, 497.6 18.0, 167.7 54.0, 797.4 57.0, 497.1 19.0, 167.5 57.0, 798.6 60.0, 498.8 20.0, 168.8 60.0, 802.1 ``` Master ``` 5.0, 743.4 10.0, 665.7 15.0, 702.3 20.0, 727.5 25.0, 740.7 30.0, 757.5 35.0, 760.3 40.0, 768.5 45.0, 775.7 50.0, 776.8 55.0, 781.1 60.0, 786.5 65.0, 786.8 70.0, 790.1 75.0, 789.7 80.0, 789.1 85.0, 793.2 90.0, 793.8 95.0, 795.9 100.0, 796.0 3.0, 383.1 1.0, 129.0 3.0, 337.0 6.0, 445.0 2.0, 149.6 6.0, 670.6 9.0, 445.3 3.0, 159.6 9.0, 678.6 12.0, 474.9 4.0, 164.1 12.0, 705.5 15.0, 480.8 5.0, 167.2 15.0, 718.3 18.0, 490.3 6.0, 169.1 18.0, 733.3 21.0, 493.9 7.0, 168.5 21.0, 742.5 24.0, 503.8 8.0, 171.9 24.0, 756.4 27.0, 506.7 9.0, 171.3 27.0, 759.8 30.0, 508.7 10.0, 172.4 30.0, 767.1 33.0, 515.7 11.0, 174.2 33.0, 773.7 36.0, 516.7 12.0, 170.4 36.0, 781.7 39.0, 519.1 13.0, 174.4 39.0, 782.1 42.0, 515.7 14.0, 174.1 42.0, 787.0 45.0, 519.2 15.0, 172.7 45.0, 788.1 48.0, 522.2 16.0, 175.4 48.0, 791.7 51.0, 519.6 17.0, 175.1 51.0, 795.7 54.0, 518.5 18.0, 174.8 54.0, 795.8 57.0, 519.1 19.0, 174.4 57.0, 796.6 60.0, 521.5 20.0, 175.6 60.0, 800.1 ``` <details> <summary>Benchmarking script </summary> ``` import torch from matplotlib import pyplot as plt from torch.utils.benchmark import Timer, Compare import math import click print(torch.cuda.get_device_capability()) # check that we are on Volta (compute capability 7,0) #torch.cuda.set_device(1) # don't benchmark on anything too small, you'll see only overhead click.command() click.option('--op_str', default="torch.gt") click.option('--dtype_str', default="float", type=click.Choice(['float', 'half'])) def bench(op_str, dtype_str): if dtype_str == "float": dtype = torch.float elif dtype_str == "half": dtype = torch.half MB = 1024 * 1024 size = MB results = [] sizes = [] for _ in range(20): torch.cuda.memory.empty_cache() a=torch.randn(int(size), device="cuda", dtype=dtype) b=torch.randn(int(size), device="cuda", dtype=dtype) t = Timer(stmt=f"{op_str}(a,b)", label = op_str, sub_label=f"{size/MB} MB", description="contiguous", globals = {"a":a, "b":b}) res = t.blocked_autorange() results.append(res) sizes.append(size) size += MB del a #to save memory for next iterations del b c=Compare(results) #print(c) bw=[] bytes=[] element_size = torch.tensor([], dtype=dtype).element_size() output_element_size = 1 for res, size in zip(results,sizes): bytes_io = 2*size*element_size + output_element_size * size bytes.append(bytes_io/MB) # we'll report bandwidth in GB/s bw.append(bytes_io/res.median * 1e-9) print(f"{bytes_io/MB:7.1f}, {bw[-1]:7.1f}") sizes = [] results = [[],[],[]] size = MB for _ in range(20): torch.cuda.memory.empty_cache() M = math.floor(math.sqrt(size)) a=torch.randn(1, M, device="cuda", dtype=dtype) b=torch.randn(M, M, device="cuda", dtype=dtype) b1 = torch.randn(M, 1, device="cuda", dtype=dtype) tb = Timer(stmt=f"{op_str}(a,b)", label = op_str, sub_label=f"{M*M/MB} MB", description="MMM1", globals = {"a":a, "b":b}) t1 = Timer(stmt=f"{op_str}(a,b1)", label = op_str, sub_label=f"{M*M/MB} MB", description="M11M", globals = {"a":a, "b1":b1}) ts = Timer(stmt=f"{op_str}(b,1.)", label = op_str, sub_label=f"{M*M/MB} MB", description="scalar", globals = {"a":a, "b":b}) res = [t.blocked_autorange() for t in (tb, t1, ts)] for (rl, r) in zip(results, res): rl.append(r) sizes.append(M) size += MB del a #to save memory for next iterations del b comps = [Compare(r) for r in results] #[print(c) for c in comps] bw=[[],[],[]] for res, res1, ress, size in zip(results[0],results[1],results[2], sizes): bytes_io = (size+size*size)*element_size + output_element_size * size*size #(size+size+size*size)*4 bytes_io1 = (size+size)*element_size + output_element_size * size*size #(size+size+size*size)*4 bytes_ios = (size*size)*element_size + output_element_size * size * size bytes_iol = (bytes_io, bytes_io1, bytes_ios) for (bw_elem, bytes_elem, res_elem) in zip(bw, bytes_iol, (res, res1, ress)): bw_elem.append(bytes_elem/res_elem.median * 1e-9) print(f"{bytes_iol[0]/MB:7.1f}, {bw[0][-1]:7.1f}", f"{bytes_iol[1]/MB:7.1f}, {bw[1][-1]:7.1f}", f"{bytes_iol[2]/MB:7.1f}, {bw[2][-1]:7.1f}") if __name__ == '__main__': bench() ``` </details> Pull Request resolved: https://github.com/pytorch/pytorch/pull/69111 Reviewed By: mruberry Differential Revision: D32851098 Pulled By: ngimel fbshipit-source-id: cfb83922b2e8eb6a0ad0621ff07c2dada9c8e626