llama-cpp-sys-4 0.10.0

Low Level Bindings to llama.cpp
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

#ifndef HTP_SOFTMAX_OPS_H
#define HTP_SOFTMAX_OPS_H

#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>
#include <math.h>
#include "hex-fastdiv.h"
#include "hex-common.h"
#include "htp-vtcm.h"

enum htp_softmax_kernel_type {
    HTP_SOFTMAX_KERNEL_NOMASK = 0,
    HTP_SOFTMAX_KERNEL_MASK_F32,
    HTP_SOFTMAX_KERNEL_MASK_F16,
    HTP_SOFTMAX_KERNEL_COUNT,
};

struct htp_softmax_kernel_params {
    uint8_t  kernel_type;
    uint32_t n_threads;
    uint32_t vtcm_size;

    uint32_t n_head;
    uint32_t n_head_log2;
    uint32_t use_src1;
    uint32_t use_f16;

    float    scale;
    float    max_bias;
    float    m0;
    float    m1;

    struct fastdiv_values div_ne01;
    struct fastdiv_values div_ne02;
    struct fastdiv_values div_ne12;
    struct fastdiv_values div_ne13;
};

#if defined(__cplusplus)
static_assert(sizeof(struct htp_softmax_kernel_params) <= 128, "htp_softmax_kernel_params is too large for kernel_params blob");
#else
_Static_assert(sizeof(struct htp_softmax_kernel_params) <= 128, "htp_softmax_kernel_params is too large for kernel_params blob");
#endif

struct htp_softmax_vtcm_layout {
    size_t total_bytes;
    size_t off_src0;
    size_t off_dst;
    size_t off_src1;

    size_t src0_bytes_per_thread;
    size_t dst_bytes_per_thread;
    size_t src1_bytes_per_thread;

    size_t src0_spad_half_size;
    size_t dst_spad_half_size;
    size_t src1_spad_half_size;
};

static inline void htp_softmax_vtcm_layout_build(
    struct htp_softmax_vtcm_layout * L,
    uint32_t ne00,
    uint32_t ne10,
    bool use_src1,
    bool use_f16,
    uint32_t n_threads
) {
    size_t src0_row_size = ne00 * sizeof(float);
    size_t dst_row_size  = ne00 * sizeof(float);
    size_t src1_row_size = use_src1 ? (ne10 * (use_f16 ? 2 : 4)) : 0;

    size_t src0_row_size_aligned = hex_round_up(src0_row_size, 128);
    size_t dst_row_size_aligned  = hex_round_up(dst_row_size,  128);
    size_t src1_row_size_aligned = use_src1 ? hex_round_up(src1_row_size, 128) : 0;

    L->src0_spad_half_size = src0_row_size_aligned;
    L->dst_spad_half_size  = dst_row_size_aligned;
    L->src1_spad_half_size = src1_row_size_aligned;

    // Double buffering: 2 half-buffers per thread
    L->src0_bytes_per_thread = src0_row_size_aligned * 2;
    L->dst_bytes_per_thread  = dst_row_size_aligned  * 2;
    L->src1_bytes_per_thread = src1_row_size_aligned * 2;

    size_t off = 0;
    VTCM_LAYOUT_ALLOC(off, off_src0, L->src0_bytes_per_thread * n_threads);
    VTCM_LAYOUT_ALLOC(off, off_dst,  L->dst_bytes_per_thread  * n_threads);
    VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_src1, L->src1_bytes_per_thread * n_threads, use_src1);

    L->total_bytes = off;
}

#endif // HTP_SOFTMAX_OPS_H