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

use super::Context;
use super::Operation;
use crate::ffi::*;
use crate::{Error, API};

impl API {
    /// Performs a general matrix-matrix multiplication.
    ///
    /// Note: the matrices are expected to be ordered column-major (FORTRAN-style).
    #[allow(clippy::many_single_char_names)]
    #[allow(clippy::too_many_arguments)]
    pub fn gemm(
        context: &Context,
        transa: Operation,
        transb: Operation,
        m: i32,
        n: i32,
        k: i32,
        alpha: *mut f32,
        a: *mut f32,
        lda: i32,
        b: *mut f32,
        ldb: i32,
        beta: *mut f32,
        c: *mut f32,
        ldc: i32,
    ) -> Result<(), Error> {
        unsafe {
            Self::ffi_sgemm(
                *context.id_c(),
                transa.as_c(),
                transb.as_c(),
                m,
                n,
                k,
                alpha,
                a,
                lda,
                b,
                ldb,
                beta,
                c,
                ldc,
            )
        }
    }

    /// Note: the matrices are expected to be ordered column-major (FORTRAN-style).
    #[allow(clippy::many_single_char_names)]
    #[allow(clippy::too_many_arguments)]
    unsafe fn ffi_sgemm(
        handle: cublasHandle_t,
        transa: cublasOperation_t,
        transb: cublasOperation_t,
        m: i32,
        n: i32,
        k: i32,
        alpha: *mut f32,
        a: *mut f32,
        lda: i32,
        b: *mut f32,
        ldb: i32,
        beta: *mut f32,
        c: *mut f32,
        ldc: i32,
    ) -> Result<(), Error> {
        match cublasSgemm_v2(
            handle, transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
        ) {
            cublasStatus_t::CUBLAS_STATUS_SUCCESS => Ok(()),
            cublasStatus_t::CUBLAS_STATUS_NOT_INITIALIZED => Err(Error::NotInitialized),
            cublasStatus_t::CUBLAS_STATUS_INVALID_VALUE => {
                Err(Error::InvalidValue("m, n, or k < 0"))
            }
            cublasStatus_t::CUBLAS_STATUS_ARCH_MISMATCH => Err(Error::ArchMismatch),
            cublasStatus_t::CUBLAS_STATUS_EXECUTION_FAILED => Err(Error::ExecutionFailed),
            status => Err(Error::Unknown(
                "Unable to calculate axpy (alpha * x + y).",
                status as i32 as u64,
            )),
        }
    }
}

#[cfg(test)]
mod test {
    use crate::api::context::Context;
    use crate::api::enums::PointerMode;
    use crate::chore::*;
    use crate::co::tensor::SharedTensor;
    use crate::ffi::*;
    use crate::API;

    #[test]
    fn use_cuda_memory_for_gemm() {
        test_setup();

        let native = get_native_backend();
        let cuda = get_cuda_backend();

        // set up alpha
        let alpha = filled_tensor(&native, 1, 1f32);

        // set up beta
        let beta = filled_tensor(&native, 1, 0f32);

        // set up a
        let mut a = SharedTensor::<f32>::new(&vec![3, 2]);
        write_to_memory(
            a.write_only(native.device()).unwrap(),
            &[2f32, 5f32, 2f32, 5f32, 2f32, 5f32],
        );

        // set up b
        let mut b = SharedTensor::<f32>::new(&vec![2, 3]);
        write_to_memory(
            b.write_only(native.device()).unwrap(),
            &[4f32, 1f32, 1f32, 4f32, 1f32, 1f32],
        );

        // set up c
        let mut c = SharedTensor::<f32>::new(&vec![3, 3]);

        {
            let transa = cublasOperation_t::CUBLAS_OP_N;
            let transb = cublasOperation_t::CUBLAS_OP_N;
            let m = 3;
            let n = 3;
            let k = 2;
            let lda = 2;
            let ldb = 3;
            let ldc = 3;
            let cuda_mem_alpha = alpha.read(cuda.device()).unwrap();
            let cuda_mem_beta = beta.read(cuda.device()).unwrap();
            let cuda_mem_a = a.read(cuda.device()).unwrap();
            let cuda_mem_b = b.read(cuda.device()).unwrap();
            let cuda_mem_c = c.write_only(cuda.device()).unwrap();
            let mut ctx = Context::new().unwrap();
            ctx.set_pointer_mode(PointerMode::Device).unwrap();
            unsafe {
                let alpha_addr = ::std::mem::transmute::<u64, *mut f32>(*cuda_mem_alpha.id_c());
                let beta_addr = ::std::mem::transmute::<u64, *mut f32>(*cuda_mem_beta.id_c());
                let a_addr = ::std::mem::transmute::<u64, *mut f32>(*cuda_mem_a.id_c());
                let b_addr = ::std::mem::transmute::<u64, *mut f32>(*cuda_mem_b.id_c());
                let c_addr = ::std::mem::transmute::<u64, *mut f32>(*cuda_mem_c.id_c());
                API::ffi_sgemm(
                    *ctx.id_c(),
                    transa,
                    transb,
                    m,
                    n,
                    k,
                    alpha_addr,
                    b_addr,
                    ldb,
                    a_addr,
                    lda,
                    beta_addr,
                    c_addr,
                    ldc,
                )
                .unwrap();
            }
        }

        let native_c = c.read(native.device()).unwrap();
        assert_eq!(
            &[28f32, 7f32, 7f32, 28f32, 7f32, 7f32, 28f32, 7f32, 7f32],
            native_c.as_slice::<f32>()
        );

        test_teardown();
    }
}