amdgpu-conv-asm/conv01.cpp at master · aditya4d/amdgpu-conv-asm · GitHub

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
#include <iostream>
#include <hip/hip_runtime.h>
#include <hip/hip_runtime_api.h>

#include <time.h>
#include <sys/time.h>
#define USECPSEC 1000000ULL

unsigned long long dtime_usec(unsigned long long start) {
    timeval tv;
    gettimeofday(&tv, 0);
    return ((tv.tv_sec*USECPSEC) + tv.tv_usec) - start;
}

constexpr int knum_iter = 1024;

typedef float float4_t __attribute__((ext_vector_type(4)));

template<int N, int C, int H, int W, int K>
__global__ void Kernel(float4_t* output, float4_t* input, float* filter) {
    int tx = threadIdx.x;
    int bx = blockIdx.x;
    int by = blockIdx.y;
    // H * W = workitems
    // K = workgroups

    for(int n = 0; n < N; n++) {
        float4_t out = {0};
        int c = 0;
        float4_t a[4], fil_a, b[4], fil_b;
        a[0] = input[n * C * H * W + c * H * W + tx];
        a[1] = input[n * C * H * W + (c+1) * H * W + tx];
        a[2] = input[n * C * H * W + (c+2) * H * W + tx];
        a[3] = input[n * C * H * W + (c+3) * H * W + tx];
        fil_a = reinterpret_cast<float4_t*>(filter)[by * C + c];
        c+= 4;
        for(; c < C; c+=8) {

            b[0] = input[n * C * H * W + c * H * W + tx];
            b[1] = input[n * C * H * W + (c+1) * H * W + tx];
            b[2] = input[n * C * H * W + (c+2) * H * W + tx];
            b[3] = input[n * C * H * W + (c+3) * H * W + tx];

            fil_b = reinterpret_cast<float4_t*>(filter)[by * C + c];

            out += a[0] * fil_a.x;
            out += a[1] * fil_a.y;
            out += a[2] * fil_a.z;
            out += a[3] * fil_a.w;

            a[0] = input[n * C * H * W + (c+4) * H * W + tx];
            a[1] = input[n * C * H * W + (c+5) * H * W + tx];
            a[2] = input[n * C * H * W + (c+6) * H * W + tx];
            a[3] = input[n * C * H * W + (c+7) * H * W + tx];

            fil_a = reinterpret_cast<float4_t*>(filter)[by * C + c];

            out += b[0] * fil_b.x;
            out += b[1] * fil_b.y;
            out += b[2] * fil_b.z;
            out += b[3] * fil_b.w;
        }
        b[0] = input[n * C * H * W + c * H * W + tx];
        b[1] = input[n * C * H * W + (c+1) * H * W + tx];
        b[2] = input[n * C * H * W + (c+2) * H * W + tx];
        b[3] = input[n * C * H * W + (c+3) * H * W + tx];

        fil_b = reinterpret_cast<float4_t*>(filter)[by * C + c];

        out += a[0] * fil_a.x;
        out += a[1] * fil_a.y;
        out += a[2] * fil_a.z;
        out += a[3] * fil_a.w;

        out += b[0] * fil_b.x;
        out += b[1] * fil_b.y;
        out += b[2] * fil_b.z;
        out += b[3] * fil_b.w;

/*
            a[0] = input[n * C * H * W + c * H * W + tx];
            a[1] = input[n * C * H * W + (c + 1) * H * W + tx];
            a[2] = input[n * C * H * W + (c + 2) * H * W + tx];
            a[3] = input[n * C * H * W + (c + 3) * H * W + tx];


    fil = *(reinterpret_cast<float4_t*>(filter) + by * C + c);

            asm volatile("\n\
    flat_load_dwordx4 %0, %4\n\
    flat_load_dwordx4 %1, %5\n\
    flat_load_dwordx4 %2, %6\n\
    flat_load_dwordx4 %3, %7\n\
    "
    :"=v"(a[0]), "=v"(a[1]), "=v"(a[2]), "=v"(a[3])
    :"v"(input + (n*C*H*W+c*H*W+tx)), "v"(input + (n*C*H*W+(c+1)*H*W+tx)), "v"(input+(n*C*H*W+(c+2)*H*W)), "v"(input+(n*C*H*W+(c+3)*H*W))
    );

    asm volatile("\n\
    s_waitcnt vmcnt(3)\n\
    v_mad_f32 %0, %4, %8, %0\n\
    v_mad_f32 %1, %5, %8, %1\n\
    v_mad_f32 %2, %6, %8, %2\n\
    v_mad_f32 %3, %7, %8, %3\n\
    ":"=v"(out.x), "=v"(out.y), "=v"(out.z), "=v"(out.w)
    :"v"(a[0].x), "v"(a[0].y), "v"(a[0].z), "v"(a[0].w),"v"(fil.x)
    );
    asm volatile("\n\
    s_waitcnt vmcnt(2)\n\
    v_mad_f32 %0, %4, %8, %0\n\
    v_mad_f32 %1, %5, %8, %1\n\
    v_mad_f32 %2, %6, %8, %2\n\
    v_mad_f32 %3, %7, %8, %3\n\
    ":"=v"(out.x), "=v"(out.y), "=v"(out.z), "=v"(out.w)
    :"v"(a[1].x), "v"(a[1].y), "v"(a[1].z), "v"(a[1].w),"v"(fil.x)
    );
    asm volatile("\n\
    s_waitcnt vmcnt(1)\n\
    v_mad_f32 %0, %4, %8, %0\n\
    v_mad_f32 %1, %5, %8, %1\n\
    v_mad_f32 %2, %6, %8, %2\n\
    v_mad_f32 %3, %7, %8, %3\n\
    ":"=v"(out.x), "=v"(out.y), "=v"(out.z), "=v"(out.w)
    :"v"(a[2].x), "v"(a[2].y), "v"(a[2].z), "v"(a[2].w),"v"(fil.x)
    );
    asm volatile("\n\
    s_waitcnt vmcnt(0)\n\
    v_mad_f32 %0, %4, %8, %0\n\
    v_mad_f32 %1, %5, %8, %1\n\
    v_mad_f32 %2, %6, %8, %2\n\
    v_mad_f32 %3, %7, %8, %3\n\
    ":"=v"(out.x), "=v"(out.y), "=v"(out.z), "=v"(out.w)
    :"v"(a[3].x), "v"(a[3].y), "v"(a[3].z), "v"(a[3].w),"v"(fil.x)
    );
*/
        output[n * K * H * W + by * H * W + tx] = out;
    }
}

int main() {
    constexpr int N = 64, C = 64, H = 56, W = 56, K = 64;
    float4_t *output, *input;
    float *filter;
    hipMalloc(&output, sizeof(float) * N * K * H * W);
    hipMalloc(&input, sizeof(float) * N * C * H * W);
    hipMalloc(&filter, sizeof(float) * K * C);

    hipEvent_t start, stop;
    hipEventCreate(&start);
    hipEventCreate(&stop);

    hipEventRecord(start);
    hipLaunchKernelGGL((Kernel<N,C,H,W/4,K>), dim3(1,K,1), dim3(W*H/4,1,1), 0, 0, output, input,filter);
    hipEventRecord(stop);

    hipEventSynchronize(stop);
    float mss = 0.0f;
    hipEventElapsedTime(&mss, start, stop);
    std::cout << double(mss) << std::endl;
    std::cout << (double(N * C * H * W * sizeof(float)))/ double(mss * 1024 * 1024 * 1024) << std::endl;
    std::cout << (double(N * C * K * 2)) / double(mss * 1024 * 1024 * 1024) << std::endl;
}