transcribe-cpp-sys 0.4.1

Native FFI bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml
Documentation 
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

#pragma OPENCL EXTENSION cl_khr_fp16 : enable

#ifdef cl_intel_subgroups
#pragma OPENCL EXTENSION cl_intel_subgroups : enable
#else
#pragma OPENCL EXTENSION cl_khr_subgroups : enable
#endif

#ifdef cl_intel_required_subgroup_size
#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
#define INTEL_GPU 1
#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
#elif defined(cl_qcom_reqd_sub_group_size)
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
#endif

#define QK8_0 32
typedef struct {
    half d;       // delta
    char qs[QK8_0]; // quants
} block_q8_0;

#define NB_Q8_0 8

#ifdef INTEL_GPU
#define N_R0_Q8_0 4 // number of rows each subgroup works on
#define N_SG_Q8_0 2 // number of subgroups in a work group
#define N_SIMDWIDTH 16 // subgroup size
#elif defined (ADRENO_GPU)
#define N_R0_Q8_0 4
#define N_SG_Q8_0 2
#define N_SIMDWIDTH 64
#endif

#ifdef INTEL_GPU
REQD_SUBGROUP_SIZE_16
#elif defined (ADRENO_GPU)
REQD_SUBGROUP_SIZE_64
#endif
kernel void kernel_mul_mv_id_q8_0_f32(
    global char * src0,
    ulong         offset0,
    global char * src1,
    ulong         offset1,
    global char * src2,
    ulong         offset2,
    global char * dst,
    ulong         offsetd,
    int           ne00,
    int           ne01,
    ulong         nb01,
    ulong         nb02,
    int           ne11,
    int           ne12,
    ulong         nb11,
    ulong         nb12,
    int           ne20,
    int           ne21,
    ulong         nb21,
    int           ne0,
    int           ne1
) {
    src0 = (global char *)((global char *)src0 + offset0);
    src1 = (global char *)((global char *)src1 + offset1);
    src2 = (global char *)((global char *)src2 + offset2);
    dst  = (global char *)((global char *)dst  + offsetd);

    int iid1 = get_group_id(2)/ne20;
    int idx  = get_group_id(2)%ne20;

    int i02 = ((global int *) (src2 + iid1*nb21))[idx];

    int i11_ = idx % ne11;
    int i12_ = iid1;

    int i1 = idx;
    int i2 = i12_;

    global char * src0_cur = src0 + i02*nb02;
    global char * src1_cur = src1 + i11_*nb11 + i12_*nb12;

    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);

    int nb = ne00/QK8_0;

    int r0 = get_group_id(0);
    int r1 = get_group_id(1);

    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;

    ulong offset_src1 = r1*nb11;
    global float * y  = (global float *) (src1_cur + offset_src1);

    // pointers to src0 rows
    global block_q8_0 * ax[N_R0_Q8_0];
    for (int row = 0; row < N_R0_Q8_0; ++row) {
        ulong offset_src0 = (first_row + row)*nb01;
        ax[row] = (global block_q8_0 *) ((global char *) src0_cur + offset_src0);
    }

    float yl[NB_Q8_0];
    float sumf[N_R0_Q8_0] = { 0.f };

    const short ix = get_sub_group_local_id()/4;
    const short il = get_sub_group_local_id()%4;

    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;

    // each thread handles NB_Q8_0 quants at a time
    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
        for (short i = 0; i < NB_Q8_0; ++i) {
            yl[i] = yb[i];
        }

        for (short row = 0; row < N_R0_Q8_0; row++) {
            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
            float sumq = 0.f;
            for (short iq = 0; iq < NB_Q8_0; ++iq) {
                sumq += qs[iq] * yl[iq];
            }
            sumf[row] += sumq*ax[row][ib].d;
        }

        yb += N_SIMDWIDTH*NB_Q8_0;
    }

    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;

    for (int row = 0; row < N_R0_Q8_0; ++row) {
        float tot = sub_group_reduce_add(sumf[row]);

        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
            dst_f32[first_row + row] = tot;
        }
    }
}