File size: 8,511 Bytes
b3a3b15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import torch
import torch.nn as nn
try:
    import AmpSelGemm
except ImportError:
    AmpSelGemm = None
    
from HybridTensor.triton.gather_gemm_col import gather_matmul_col
from HybridTensor.triton.gather_gemm_row import gather_matmul_row
from HybridTensor.utils.utils import arg_parser, sparse_index, create_results_directory 
from HybridTensor.utils.profiling import benchmark_mlp_fwd, generate_index_sizes, save_results_to_csv, plot_results

import pandas as pd
import matplotlib.pyplot as plt

from tqdm import tqdm  # For progress bars

# implement standard MLP block with ReLU activation
class StandardMLPBlock(nn.Module):
    def __init__(self, in_features, hidden_features=None, bias=False, device='cuda'):
        super(StandardMLPBlock, self).__init__()
        hidden_features = hidden_features or in_features*4
        
        # this is stored in correct order; don't need to transpose for the CUTLASS kernel
        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias, device=device, dtype=torch.float16)
        
        # this is stored in row major; need to transpose for the CUTLASS kernel
        self.fc2 = nn.Linear(hidden_features, in_features, bias=bias, device=device, dtype=torch.float16)
        self.relu = nn.ReLU()

    def forward(self, x):
        # fc1 : d x (4d)
        out = self.fc1(x)
        # B x (4d)
        out = self.relu(out)
        
        #fc2: (4d) x d
        out = self.fc2(out)
        return out
    
    def sparsify(self, zero_index):
        self.fc1.weight.data[zero_index, :] = 0.0
        self.fc2.weight.data[:, zero_index] = 0.0
        
class SelectiveMLP(nn.Module):
    def __init__(self, in_features, hidden_features=None, bias=False, requires_grad = False, device='cuda', dtype=torch.float16, activation='relu'):
        super(SelectiveMLP, self).__init__()
        if hidden_features is None:
            hidden_features = in_features*4
            
        factory_kwargs = {'device': torch.device(device), 'dtype': dtype}
        self.fc1_w = torch.empty((hidden_features, in_features), requires_grad=requires_grad, **factory_kwargs)
        self.fc2_w = torch.empty((hidden_features, in_features), requires_grad=requires_grad, **factory_kwargs)
        self.act = nn.ReLU()
        
    def forward(self, x, index_vec):
        index_size = index_vec.size(0)
        #AmpSelGemm.run(A=A, B=B_col_major, index_vec= index_vec, M= M, N=index_size, K=K, index_size=index_size)
        out = AmpSelGemm.run_col(A=x, B=self.fc1_w, index_vec= index_vec, M = x.size(0), N = index_size, K = self.fc1_w.size(1), index_size=index_size)
        out = self.act(out)   # need to fuse this with fc1 in the next iteration
        out = AmpSelGemm.run_row1(A=out, B=self.fc2_w, index_vec= index_vec, M = x.size(0), N = self.fc2_w.size(1), K = index_size, index_size=index_size)
        return out
    
    def load_from_MLP(self, mlp):
        self.fc1_w = mlp.fc1.weight  
        self.fc2_w = mlp.fc2.weight.t().contiguous()  
        return self
    
class SelectiveMLPTriton(SelectiveMLP):
    def __init__(self, in_features, hidden_features=None, bias=False, requires_grad = False, device='cuda', dtype=torch.float16, activation='relu'):
        super(SelectiveMLPTriton, self).__init__(in_features, hidden_features, bias, requires_grad, device, dtype, activation)
        self.activation = activation

    def forward(self, x, index_vec):
        out = gather_matmul_col(x, self.fc1_w, index_vec, activations=self.activation)
        out = gather_matmul_row(out, self.fc2_w, index_vec)
        return out

def SelectiveMLPFunc(x, fc1_w, fc2_w, index_vec, bias1 = None, bias2 = None, activation='relu'):
    out = gather_matmul_col(x, fc1_w, index_vec, bias = bias1, activations=activation)
    out = gather_matmul_row(out, fc2_w, index_vec, bias = bias2)
    return out

def profile_mlps(args, index_size):
    # Create standard MLP block
    standardmlp = StandardMLPBlock(args.in_features, args.hidden_features)
    if AmpSelGemm is not None:
        selectiveMLP = SelectiveMLP(args.in_features, args.hidden_features).load_from_MLP(standardmlp)
    selectiveMLPTriton = SelectiveMLPTriton(args.in_features, args.hidden_features).load_from_MLP(standardmlp)
    
    # Test input
    x = torch.randn(args.batch_size, args.in_features, dtype=torch.float16, device='cuda')
    index_vec, zero_index = sparse_index(index_size, args.hidden_features or args.in_features*4)
    standardmlp.sparsify(zero_index)
    
    # Measure execution time
    std_out, standardmlp_time = benchmark_mlp_fwd(x, standardmlp, index_vec=None, iterations=args.iterations, print_result=False)
    if AmpSelGemm is not None:
        cutlass_out, selectiveMLP_time = benchmark_mlp_fwd(x, selectiveMLP, index_vec=index_vec, iterations=args.iterations, print_result=False)
    triton_out, selectiveMLPTriton_time = benchmark_mlp_fwd(x, selectiveMLPTriton, index_vec=index_vec, iterations=args.iterations, print_result=False)
    
    # Optionally check results
    if args.check_results:
        print('Standard MLP output:', std_out[0])
        if AmpSelGemm is not None:
            print('Selective MLP Cutlass output:', cutlass_out)
        print('Selective MLP Triton output:', triton_out)
    
    # Calculate speedups
    triton_speedup = standardmlp_time / selectiveMLPTriton_time if selectiveMLPTriton_time > 0 else float('inf')
    if AmpSelGemm is not None:
        cutlass_speedup = standardmlp_time / selectiveMLP_time if selectiveMLP_time > 0 else float('inf')
    
    return {
        'index_size': index_size,
        'standard_time': standardmlp_time,
        'selective_cutlass_time': selectiveMLP_time,
        'selective_triton_time': selectiveMLPTriton_time,
        'cutlass_speedup': cutlass_speedup,
        'triton_speedup': triton_speedup
    }
    
def run_profiling_over_index_sizes(args, index_sizes):
    results = []
    for size in tqdm(index_sizes, desc="Profiling MLPs"):
        result = profile_mlps(args, size)
        results.append(result)
    return pd.DataFrame(results)


'''
if __name__ == '__main__':
    args = arg_parser()

    index_sizes = generate_index_sizes(args.hidden_features)
    
    # create standard MLP block
    standardmlp = StandardMLPBlock(args.in_features, args.hidden_features)
    selectiveMLP = SelectiveMLP(args.in_features, args.hidden_features).load_from_MLP(standardmlp)
    selectiveMLPTriton = SelectiveMLPTriton(args.in_features, args.hidden_features).load_from_MLP(standardmlp)
    
    
    # test input
    x = torch.randn(args.batch_size, args.in_features, dtype=torch.float16, device='cuda')
    index_vec, zero_index = sparse_index(args.index_size, args.hidden_features or args.in_features*4)
    standardmlp.sparsify(zero_index)
    
    # measure execution time
    std_out, standardmlp_time = benchmark_mlp_fwd(x, standardmlp, index_vec= None, iterations=args.iterations, print_result=True)
    cutlass_out, selectiveMLP_time = benchmark_mlp_fwd(x, selectiveMLP, index_vec= index_vec, iterations=args.iterations, print_result=True)
    triton_out, selectiveMLPTriton_time = benchmark_mlp_fwd(x, selectiveMLPTriton, index_vec= index_vec, iterations=args.iterations, print_result=True)
    
    if args.check_results:
        print('Standard MLP output:', std_out[0])
        print('Selective MLP Cutlass output:', cutlass_out)
        print('Selective MLP Triton output:', triton_out)
    
    triton_speedup = standardmlp_time/selectiveMLPTriton_time
    cutlass_speedup = standardmlp_time/selectiveMLP_time
    
    print(f"Speedup of Cutlass implementation over standard MLP: {cutlass_speedup}")
    print(f"Speedup of Triton implementation over standard MLP: {triton_speedup}")

'''
    
if __name__ == '__main__':
    args = arg_parser()
    
    print(f"Profiling MLPs")
    # Results directory
    results_dir = create_results_directory(args.results_dir)
    
    # Define the range of index sizes you want to profile
    index_sizes = generate_index_sizes(args.hidden_features)
    
    # Run profiling over different index sizes
    profiling_results = run_profiling_over_index_sizes(args, index_sizes)
    
    # Save the results to a CSV file
    save_results_to_csv(profiling_results, filename_prefix='mlp_profiling_results', results_dir=results_dir)
    
    # Plot the results
    plot_results(profiling_results, output_prefix='mlp_profiling', results_dir=results_dir)
    
    # Optionally, print the DataFrame
    if args.check_results:
        print(profiling_results)