llama_cpp 0.15.4 → 0.16.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +10 -0
- data/ext/llama_cpp/extconf.rb +1 -2
- data/ext/llama_cpp/llama_cpp.cpp +15 -3
- data/lib/llama_cpp/version.rb +2 -2
- data/sig/llama_cpp.rbs +13 -1
- data/vendor/tmp/llama.cpp/Makefile +62 -35
- data/vendor/tmp/llama.cpp/ggml-alloc.c +4 -4
- data/vendor/tmp/llama.cpp/ggml-backend.c +5 -5
- data/vendor/tmp/llama.cpp/ggml-backend.h +1 -1
- data/vendor/tmp/llama.cpp/ggml-cuda/acc.cu +47 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/arange.cu +34 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/argsort.cu +103 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/binbcast.cu +280 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/clamp.cu +34 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/concat.cu +196 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/convert.cu +686 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/cpy.cu +490 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/diagmask.cu +40 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/dmmv.cu +662 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f16.cu +319 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f32.cu +312 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn.cu +345 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/getrows.cu +178 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/im2col.cu +104 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/mmq.cu +1564 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/mmvq.cu +404 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/norm.cu +221 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/pad.cu +49 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/pool2d.cu +94 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/quantize.cu +45 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/rope.cu +271 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/scale.cu +31 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/softmax.cu +205 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/sumrows.cu +40 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +5 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +10 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +9 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +10 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +10 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +8 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/tsembd.cu +47 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/unary.cu +266 -0
- data/vendor/tmp/llama.cpp/ggml-cuda/upscale.cu +51 -0
- data/vendor/tmp/llama.cpp/ggml-cuda.cu +8 -6
- data/vendor/tmp/llama.cpp/ggml-kompute.cpp +21 -6
- data/vendor/tmp/llama.cpp/ggml-metal.h +1 -1
- data/vendor/tmp/llama.cpp/ggml-metal.m +34 -24
- data/vendor/tmp/llama.cpp/ggml-metal.metal +83 -59
- data/vendor/tmp/llama.cpp/ggml-rpc.cpp +2 -2
- data/vendor/tmp/llama.cpp/ggml-sycl.cpp +7 -67
- data/vendor/tmp/llama.cpp/ggml-vulkan-shaders.hpp +99301 -39793
- data/vendor/tmp/llama.cpp/ggml-vulkan.cpp +456 -329
- data/vendor/tmp/llama.cpp/ggml.c +178 -330
- data/vendor/tmp/llama.cpp/ggml.h +9 -28
- data/vendor/tmp/llama.cpp/llama.cpp +242 -426
- data/vendor/tmp/llama.cpp/llama.h +17 -43
- metadata +121 -6
- data/vendor/tmp/llama.cpp/ggml-mpi.c +0 -216
- data/vendor/tmp/llama.cpp/ggml-mpi.h +0 -39
- data/vendor/tmp/llama.cpp/ggml-opencl.cpp +0 -2305
- data/vendor/tmp/llama.cpp/ggml-opencl.h +0 -36
|
@@ -1,2305 +0,0 @@
|
|
|
1
|
-
#include "ggml.h"
|
|
2
|
-
#include "ggml-opencl.h"
|
|
3
|
-
#include "ggml-backend-impl.h"
|
|
4
|
-
|
|
5
|
-
#include <array>
|
|
6
|
-
#include <atomic>
|
|
7
|
-
#include <cstdio>
|
|
8
|
-
#include <cstdlib>
|
|
9
|
-
#include <cstring>
|
|
10
|
-
#include <limits>
|
|
11
|
-
#include <sstream>
|
|
12
|
-
#include <vector>
|
|
13
|
-
|
|
14
|
-
#define CL_TARGET_OPENCL_VERSION 120
|
|
15
|
-
#include <clblast.h>
|
|
16
|
-
|
|
17
|
-
#if defined(_MSC_VER)
|
|
18
|
-
#pragma warning(disable: 4244 4267) // possible loss of data
|
|
19
|
-
#endif
|
|
20
|
-
|
|
21
|
-
#define CL_DMMV_LOCAL_SIZE 32
|
|
22
|
-
|
|
23
|
-
#ifndef K_QUANTS_PER_ITERATION
|
|
24
|
-
#define K_QUANTS_PER_ITERATION 1
|
|
25
|
-
#else
|
|
26
|
-
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
|
|
27
|
-
#endif
|
|
28
|
-
|
|
29
|
-
#define MULTILINE_QUOTE(...) #__VA_ARGS__
|
|
30
|
-
static std::string program_source = MULTILINE_QUOTE(
|
|
31
|
-
|
|
32
|
-
typedef char int8_t;
|
|
33
|
-
typedef uchar uint8_t;
|
|
34
|
-
typedef short int16_t;
|
|
35
|
-
typedef ushort uint16_t;
|
|
36
|
-
typedef int int32_t;
|
|
37
|
-
typedef uint uint32_t;
|
|
38
|
-
|
|
39
|
-
struct __attribute__ ((packed)) block_q4_0
|
|
40
|
-
{
|
|
41
|
-
half d;
|
|
42
|
-
uint8_t qs[QK4_0 / 2];
|
|
43
|
-
};
|
|
44
|
-
|
|
45
|
-
struct __attribute__ ((packed)) block_q4_1
|
|
46
|
-
{
|
|
47
|
-
half d;
|
|
48
|
-
half m;
|
|
49
|
-
uint8_t qs[QK4_1 / 2];
|
|
50
|
-
};
|
|
51
|
-
|
|
52
|
-
struct __attribute__ ((packed)) block_q5_0
|
|
53
|
-
{
|
|
54
|
-
half d;
|
|
55
|
-
uint32_t qh;
|
|
56
|
-
uint8_t qs[QK5_0 / 2];
|
|
57
|
-
};
|
|
58
|
-
|
|
59
|
-
struct __attribute__ ((packed)) block_q5_1
|
|
60
|
-
{
|
|
61
|
-
half d;
|
|
62
|
-
half m;
|
|
63
|
-
uint32_t qh;
|
|
64
|
-
uint8_t qs[QK5_1 / 2];
|
|
65
|
-
};
|
|
66
|
-
|
|
67
|
-
struct __attribute__ ((packed)) block_q8_0
|
|
68
|
-
{
|
|
69
|
-
half d;
|
|
70
|
-
int8_t qs[QK8_0];
|
|
71
|
-
};
|
|
72
|
-
|
|
73
|
-
struct __attribute__((packed)) block_q2_K
|
|
74
|
-
{
|
|
75
|
-
uint8_t scales[16];
|
|
76
|
-
uint8_t qs[64];
|
|
77
|
-
half d;
|
|
78
|
-
half dmin;
|
|
79
|
-
};
|
|
80
|
-
|
|
81
|
-
struct __attribute__((packed)) block_q3_K
|
|
82
|
-
{
|
|
83
|
-
uint8_t hmask[32];
|
|
84
|
-
uint8_t qs[64];
|
|
85
|
-
uint8_t scales[12];
|
|
86
|
-
half d;
|
|
87
|
-
};
|
|
88
|
-
|
|
89
|
-
struct __attribute__((packed)) block_q4_K
|
|
90
|
-
{
|
|
91
|
-
half d;
|
|
92
|
-
half dmin;
|
|
93
|
-
uint8_t scales[12];
|
|
94
|
-
uint8_t qs[128];
|
|
95
|
-
};
|
|
96
|
-
|
|
97
|
-
struct __attribute__((packed)) block_q5_K
|
|
98
|
-
{
|
|
99
|
-
half d;
|
|
100
|
-
half dmin;
|
|
101
|
-
uint8_t scales[12];
|
|
102
|
-
uint8_t qh[32];
|
|
103
|
-
uint8_t qs[128];
|
|
104
|
-
};
|
|
105
|
-
|
|
106
|
-
struct __attribute__((packed)) block_q6_K
|
|
107
|
-
{
|
|
108
|
-
uint8_t ql[128];
|
|
109
|
-
uint8_t qh[64];
|
|
110
|
-
int8_t scales[16];
|
|
111
|
-
half d;
|
|
112
|
-
};
|
|
113
|
-
|
|
114
|
-
__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
|
|
115
|
-
const uint i = get_global_id(0);
|
|
116
|
-
|
|
117
|
-
y[i] = vload_half(0, &x[i]);
|
|
118
|
-
}
|
|
119
|
-
|
|
120
|
-
void dequantize_q4_0(__global const struct block_q4_0* x, const int ib, const int iqs, float* v0, float* v1) {
|
|
121
|
-
const float d = vload_half(0, &x[ib].d);
|
|
122
|
-
|
|
123
|
-
const uint8_t vui = x[ib].qs[iqs];
|
|
124
|
-
|
|
125
|
-
const int8_t vi0 = vui & 0xF;
|
|
126
|
-
const int8_t vi1 = vui >> 4;
|
|
127
|
-
|
|
128
|
-
*v0 = (vi0 - 8)*d;
|
|
129
|
-
*v1 = (vi1 - 8)*d;
|
|
130
|
-
}
|
|
131
|
-
void dequantize_q4_1(__global const struct block_q4_1* x, const int ib, const int iqs, float* v0, float* v1) {
|
|
132
|
-
const float d = vload_half(0, &x[ib].d);
|
|
133
|
-
const float m = vload_half(0, &x[ib].m);
|
|
134
|
-
|
|
135
|
-
const uint8_t vui = x[ib].qs[iqs];
|
|
136
|
-
|
|
137
|
-
const int8_t vi0 = vui & 0xF;
|
|
138
|
-
const int8_t vi1 = vui >> 4;
|
|
139
|
-
|
|
140
|
-
*v0 = vi0*d + m;
|
|
141
|
-
*v1 = vi1*d + m;
|
|
142
|
-
}
|
|
143
|
-
void dequantize_q5_0(__global const struct block_q5_0* x, const int ib, const int iqs, float* v0, float* v1) {
|
|
144
|
-
const float d = vload_half(0, &x[ib].d);
|
|
145
|
-
|
|
146
|
-
uint32_t qh = x[ib].qh;
|
|
147
|
-
|
|
148
|
-
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
149
|
-
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
|
150
|
-
|
|
151
|
-
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
|
|
152
|
-
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16;
|
|
153
|
-
|
|
154
|
-
*v0 = x0*d;
|
|
155
|
-
*v1 = x1*d;
|
|
156
|
-
}
|
|
157
|
-
void dequantize_q5_1(__global const struct block_q5_1* x, const int ib, const int iqs, float* v0, float* v1) {
|
|
158
|
-
const float d = vload_half(0, &x[ib].d);
|
|
159
|
-
const float m = vload_half(0, &x[ib].m);
|
|
160
|
-
|
|
161
|
-
uint32_t qh = x[ib].qh;
|
|
162
|
-
|
|
163
|
-
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
164
|
-
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
|
165
|
-
|
|
166
|
-
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
|
|
167
|
-
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1);
|
|
168
|
-
|
|
169
|
-
*v0 = x0*d + m;
|
|
170
|
-
*v1 = x1*d + m;
|
|
171
|
-
}
|
|
172
|
-
void dequantize_q8_0(__global const struct block_q8_0* x, const int ib, const int iqs, float* v0, float* v1) {
|
|
173
|
-
const float d = vload_half(0, &x[ib].d);
|
|
174
|
-
|
|
175
|
-
const int8_t vi0 = x[ib].qs[iqs + 0];
|
|
176
|
-
const int8_t vi1 = x[ib].qs[iqs + 1];
|
|
177
|
-
|
|
178
|
-
*v0 = vi0*d;
|
|
179
|
-
*v1 = vi1*d;
|
|
180
|
-
}
|
|
181
|
-
void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float* v1){
|
|
182
|
-
*v0 = vload_half(0, &x[ib + 0]);
|
|
183
|
-
*v1 = vload_half(0, &x[ib + 1]);
|
|
184
|
-
}
|
|
185
|
-
);
|
|
186
|
-
|
|
187
|
-
static std::string k_quants_source = MULTILINE_QUOTE(
|
|
188
|
-
inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
|
|
189
|
-
{
|
|
190
|
-
if (j < 4)
|
|
191
|
-
{
|
|
192
|
-
*d = q[j] & 63;
|
|
193
|
-
*m = q[j + 4] & 63;
|
|
194
|
-
}
|
|
195
|
-
else
|
|
196
|
-
{
|
|
197
|
-
*d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
|
|
198
|
-
*m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
|
|
199
|
-
}
|
|
200
|
-
}
|
|
201
|
-
|
|
202
|
-
__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
|
|
203
|
-
{
|
|
204
|
-
const int i = get_group_id(0) + get_global_offset(0);
|
|
205
|
-
const int tid = get_local_id(0);
|
|
206
|
-
const int n = tid / 32;
|
|
207
|
-
const int l = tid - 32 * n;
|
|
208
|
-
const int is = 8 * n + l / 16;
|
|
209
|
-
|
|
210
|
-
const uint8_t q = x[i].qs[32 * n + l];
|
|
211
|
-
__global float *y = yy + get_group_id(0) * QK_K + 128 * n;
|
|
212
|
-
|
|
213
|
-
const float dall = vload_half(0, &x[i].d);
|
|
214
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
215
|
-
|
|
216
|
-
y[l + 0] = dall * (x[i].scales[is + 0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is + 0] >> 4);
|
|
217
|
-
y[l + 32] = dall * (x[i].scales[is + 2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is + 2] >> 4);
|
|
218
|
-
y[l + 64] = dall * (x[i].scales[is + 4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is + 4] >> 4);
|
|
219
|
-
y[l + 96] = dall * (x[i].scales[is + 6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is + 6] >> 4);
|
|
220
|
-
}
|
|
221
|
-
|
|
222
|
-
__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
|
|
223
|
-
{
|
|
224
|
-
int r = get_local_id(0) / 4;
|
|
225
|
-
int i = get_group_id(0) + get_global_offset(0);
|
|
226
|
-
int tid = r / 2;
|
|
227
|
-
int is0 = r % 2;
|
|
228
|
-
int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
|
|
229
|
-
int n = tid / 4;
|
|
230
|
-
int j = tid - 4 * n;
|
|
231
|
-
|
|
232
|
-
uint8_t m = 1 << (4 * n + j);
|
|
233
|
-
int is = 8 * n + 2 * j + is0;
|
|
234
|
-
int shift = 2 * j;
|
|
235
|
-
|
|
236
|
-
int8_t us = is < 4 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 8] >> 0) & 3) << 4)
|
|
237
|
-
: is < 8 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 4] >> 2) & 3) << 4)
|
|
238
|
-
: is < 12 ? (x[i].scales[is - 8] >> 4) | (((x[i].scales[is + 0] >> 4) & 3) << 4)
|
|
239
|
-
: (x[i].scales[is - 8] >> 4) | (((x[i].scales[is - 4] >> 6) & 3) << 4);
|
|
240
|
-
float d_all = vload_half(0, &x[i].d);
|
|
241
|
-
float dl = d_all * (us - 32);
|
|
242
|
-
|
|
243
|
-
__global float *y = yy + get_group_id(0) * QK_K + 128 * n + 32 * j;
|
|
244
|
-
const __global uint8_t *q = x[i].qs + 32 * n;
|
|
245
|
-
const __global uint8_t *hm = x[i].hmask;
|
|
246
|
-
|
|
247
|
-
for (int l = l0; l < l0 + 4; ++l)
|
|
248
|
-
y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
|
|
249
|
-
}
|
|
250
|
-
|
|
251
|
-
__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
|
|
252
|
-
{
|
|
253
|
-
const int i = get_group_id(0) + get_global_offset(0);
|
|
254
|
-
const int tid = get_local_id(0);
|
|
255
|
-
const int il = tid / 8;
|
|
256
|
-
const int ir = tid % 8;
|
|
257
|
-
const int is = 2 * il;
|
|
258
|
-
const int n = 4;
|
|
259
|
-
|
|
260
|
-
__global float *y = yy + get_group_id(0) * QK_K + 64 * il + n * ir;
|
|
261
|
-
|
|
262
|
-
const float dall = vload_half(0, &x[i].d);
|
|
263
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
264
|
-
|
|
265
|
-
__global const uint8_t *q = x[i].qs + 32 * il + n * ir;
|
|
266
|
-
|
|
267
|
-
uint8_t sc, m;
|
|
268
|
-
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
|
|
269
|
-
float d1 = dall * sc;
|
|
270
|
-
float m1 = dmin * m;
|
|
271
|
-
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
|
|
272
|
-
float d2 = dall * sc;
|
|
273
|
-
float m2 = dmin * m;
|
|
274
|
-
for (int l = 0; l < n; ++l)
|
|
275
|
-
{
|
|
276
|
-
y[l + 0] = d1 * (q[l] & 0xF) - m1;
|
|
277
|
-
y[l + 32] = d2 * (q[l] >> 4) - m2;
|
|
278
|
-
}
|
|
279
|
-
}
|
|
280
|
-
|
|
281
|
-
__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
|
|
282
|
-
{
|
|
283
|
-
const int i = get_group_id(0) + get_global_offset(0);
|
|
284
|
-
const int tid = get_local_id(0);
|
|
285
|
-
const int il = tid / 16;
|
|
286
|
-
const int ir = tid % 16;
|
|
287
|
-
const int is = 2 * il;
|
|
288
|
-
|
|
289
|
-
__global float *y = yy + get_group_id(0) * QK_K + 64 * il + 2 * ir;
|
|
290
|
-
|
|
291
|
-
const float dall = vload_half(0, &x[i].d);
|
|
292
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
293
|
-
|
|
294
|
-
__global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
|
|
295
|
-
__global const uint8_t *qh = x[i].qh + 2 * ir;
|
|
296
|
-
|
|
297
|
-
uint8_t sc, m;
|
|
298
|
-
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
|
|
299
|
-
const float d1 = dall * sc;
|
|
300
|
-
const float m1 = dmin * m;
|
|
301
|
-
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
|
|
302
|
-
const float d2 = dall * sc;
|
|
303
|
-
const float m2 = dmin * m;
|
|
304
|
-
|
|
305
|
-
uint8_t hm = 1 << (2 * il);
|
|
306
|
-
y[0] = d1 * ((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0)) - m1;
|
|
307
|
-
y[1] = d1 * ((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0)) - m1;
|
|
308
|
-
hm <<= 1;
|
|
309
|
-
y[32] = d2 * ((ql[0] >> 4) + (qh[0] & hm ? 16 : 0)) - m2;
|
|
310
|
-
y[33] = d2 * ((ql[1] >> 4) + (qh[1] & hm ? 16 : 0)) - m2;
|
|
311
|
-
}
|
|
312
|
-
|
|
313
|
-
__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
|
|
314
|
-
{
|
|
315
|
-
const int i = get_group_id(0) + get_global_offset(0);
|
|
316
|
-
const int tid = get_local_id(0);
|
|
317
|
-
const int ip = tid / 32;
|
|
318
|
-
const int il = tid - 32 * ip;
|
|
319
|
-
const int is = 8 * ip + il / 16;
|
|
320
|
-
|
|
321
|
-
__global float *y = yy + get_group_id(0) * QK_K + 128 * ip + il;
|
|
322
|
-
|
|
323
|
-
const float d = vload_half(0, &x[i].d);
|
|
324
|
-
|
|
325
|
-
__global const uint8_t *ql = x[i].ql + 64 * ip + il;
|
|
326
|
-
const uint8_t qh = x[i].qh[32 * ip + il];
|
|
327
|
-
__global const int8_t *sc = x[i].scales + is;
|
|
328
|
-
|
|
329
|
-
y[0] = d * sc[0] * ((int8_t)((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
|
|
330
|
-
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
|
|
331
|
-
y[64] = d * sc[4] * ((int8_t)((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
|
|
332
|
-
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
|
|
333
|
-
}
|
|
334
|
-
|
|
335
|
-
__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
|
|
336
|
-
|
|
337
|
-
const int row = get_group_id(0);
|
|
338
|
-
|
|
339
|
-
const int num_blocks_per_row = ncols / QK_K;
|
|
340
|
-
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
|
|
341
|
-
|
|
342
|
-
__global const struct block_q2_K * x = xx + ib0;
|
|
343
|
-
|
|
344
|
-
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
|
|
345
|
-
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
|
|
346
|
-
|
|
347
|
-
const int step = 16/K_QUANTS_PER_ITERATION;
|
|
348
|
-
|
|
349
|
-
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
350
|
-
const int in = tid - step*im; // 0...15 or 0...7
|
|
351
|
-
|
|
352
|
-
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2
|
|
353
|
-
const int q_offset = 32*im + l0;
|
|
354
|
-
const int s_offset = 8*im;
|
|
355
|
-
const int y_offset = 128*im + l0;
|
|
356
|
-
|
|
357
|
-
tmp[16 * ix + tid] = 0;
|
|
358
|
-
|
|
359
|
-
uint32_t aux[4];
|
|
360
|
-
const uint8_t * d = (const uint8_t *)aux;
|
|
361
|
-
const uint8_t * m = (const uint8_t *)(aux + 2);
|
|
362
|
-
|
|
363
|
-
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
364
|
-
|
|
365
|
-
__global const float * y = yy + i * QK_K + y_offset;
|
|
366
|
-
__global const uint8_t * q = x[i].qs + q_offset;
|
|
367
|
-
|
|
368
|
-
const float dall = vload_half(0, &x[i].d);
|
|
369
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
370
|
-
|
|
371
|
-
__global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset);
|
|
372
|
-
aux[0] = a[0] & 0x0f0f0f0f;
|
|
373
|
-
aux[1] = a[1] & 0x0f0f0f0f;
|
|
374
|
-
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
|
|
375
|
-
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
|
|
376
|
-
|
|
377
|
-
float sum1 = 0, sum2 = 0;
|
|
378
|
-
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
|
|
379
|
-
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
|
|
380
|
-
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
|
|
381
|
-
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
|
|
382
|
-
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
|
|
383
|
-
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
|
|
384
|
-
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
|
|
385
|
-
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
|
|
386
|
-
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
|
|
387
|
-
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
|
|
388
|
-
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
|
|
389
|
-
|
|
390
|
-
}
|
|
391
|
-
tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
|
|
392
|
-
|
|
393
|
-
}
|
|
394
|
-
|
|
395
|
-
// sum up partial sums and write back result
|
|
396
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
397
|
-
for (int s=16; s>0; s>>=1) {
|
|
398
|
-
if (tid < s) {
|
|
399
|
-
tmp[tid] += tmp[tid + s];
|
|
400
|
-
}
|
|
401
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
402
|
-
}
|
|
403
|
-
if (tid == 0) {
|
|
404
|
-
dst[row] = tmp[0];
|
|
405
|
-
}
|
|
406
|
-
}
|
|
407
|
-
|
|
408
|
-
__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
|
|
409
|
-
const uint16_t kmask1 = 0x0303;
|
|
410
|
-
const uint16_t kmask2 = 0x0f0f;
|
|
411
|
-
|
|
412
|
-
const int row = get_group_id(0);
|
|
413
|
-
|
|
414
|
-
const int num_blocks_per_row = ncols / QK_K;
|
|
415
|
-
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
|
|
416
|
-
|
|
417
|
-
__global const struct block_q3_K * x = xx + ib0;
|
|
418
|
-
|
|
419
|
-
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
|
|
420
|
-
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
|
|
421
|
-
|
|
422
|
-
const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop
|
|
423
|
-
const int step = 16/K_QUANTS_PER_ITERATION;
|
|
424
|
-
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
425
|
-
const int in = tid - step*im; // 0....15 or 0...7
|
|
426
|
-
|
|
427
|
-
const uint8_t m = 1 << (4*im);
|
|
428
|
-
|
|
429
|
-
const int l0 = n*in; // 0...15 or 0...14 in steps of 2
|
|
430
|
-
const int q_offset = 32*im + l0;
|
|
431
|
-
const int y_offset = 128*im + l0;
|
|
432
|
-
|
|
433
|
-
uint16_t utmp[4];
|
|
434
|
-
const int8_t * s = (const int8_t *)utmp;
|
|
435
|
-
|
|
436
|
-
const uint16_t s_shift = 4*im;
|
|
437
|
-
|
|
438
|
-
tmp[16 * ix + tid] = 0;
|
|
439
|
-
|
|
440
|
-
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
441
|
-
|
|
442
|
-
__global const float * y = yy + i * QK_K + y_offset;
|
|
443
|
-
__global const uint8_t * q = x[i].qs + q_offset;
|
|
444
|
-
__global const uint8_t * h = x[i].hmask + l0;
|
|
445
|
-
|
|
446
|
-
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
|
|
447
|
-
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
|
|
448
|
-
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
|
|
449
|
-
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
|
|
450
|
-
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
|
|
451
|
-
|
|
452
|
-
const float d = vload_half(0, &x[i].d);
|
|
453
|
-
|
|
454
|
-
float sum = 0;
|
|
455
|
-
for (int l = 0; l < n; ++l) {
|
|
456
|
-
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
|
|
457
|
-
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
|
|
458
|
-
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
|
|
459
|
-
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
|
|
460
|
-
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
|
|
461
|
-
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
|
|
462
|
-
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
|
|
463
|
-
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
|
|
464
|
-
}
|
|
465
|
-
tmp[16 * ix + tid] += d * sum;
|
|
466
|
-
|
|
467
|
-
}
|
|
468
|
-
|
|
469
|
-
// sum up partial sums and write back result
|
|
470
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
471
|
-
for (int s=16; s>0; s>>=1) {
|
|
472
|
-
if (tid < s) {
|
|
473
|
-
tmp[tid] += tmp[tid + s];
|
|
474
|
-
}
|
|
475
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
476
|
-
}
|
|
477
|
-
if (tid == 0) {
|
|
478
|
-
dst[row] = tmp[0];
|
|
479
|
-
}
|
|
480
|
-
}
|
|
481
|
-
|
|
482
|
-
__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
|
|
483
|
-
|
|
484
|
-
//to rename it later, just to test now
|
|
485
|
-
const uint16_t kmask1 = 0x3f3f;
|
|
486
|
-
const uint16_t kmask2 = 0x0f0f;
|
|
487
|
-
const uint16_t kmask3 = 0xc0c0;
|
|
488
|
-
|
|
489
|
-
const int row = get_group_id(0);
|
|
490
|
-
const int num_blocks_per_row = ncols / QK_K;
|
|
491
|
-
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
|
|
492
|
-
|
|
493
|
-
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...15
|
|
494
|
-
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION;
|
|
495
|
-
|
|
496
|
-
const int step = 8/K_QUANTS_PER_ITERATION;
|
|
497
|
-
|
|
498
|
-
const int il = tid/step; // 0...3
|
|
499
|
-
const int ir = tid - step*il;// 0...3
|
|
500
|
-
const int n = 2*K_QUANTS_PER_ITERATION;
|
|
501
|
-
|
|
502
|
-
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
|
|
503
|
-
const int in = il%2;
|
|
504
|
-
|
|
505
|
-
const int l0 = n*(2*ir + in);
|
|
506
|
-
const int q_offset = 32*im + l0;
|
|
507
|
-
const int y_offset = 64*im + l0;
|
|
508
|
-
|
|
509
|
-
uint16_t aux[4];
|
|
510
|
-
const uint8_t * sc = (const uint8_t *)aux;
|
|
511
|
-
|
|
512
|
-
__global const struct block_q4_K * x = xx + ib0;
|
|
513
|
-
|
|
514
|
-
tmp[16 * ix + tid] = 0;
|
|
515
|
-
|
|
516
|
-
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
517
|
-
|
|
518
|
-
__global const uint8_t * q1 = x[i].qs + q_offset;
|
|
519
|
-
__global const uint8_t * q2 = q1 + 64;
|
|
520
|
-
__global const float * y1 = yy + i*QK_K + y_offset;
|
|
521
|
-
__global const float * y2 = y1 + 128;
|
|
522
|
-
|
|
523
|
-
const float dall = vload_half(0, &x[i].d);
|
|
524
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
525
|
-
|
|
526
|
-
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
|
|
527
|
-
aux[0] = a[im+0] & kmask1;
|
|
528
|
-
aux[1] = a[im+2] & kmask1;
|
|
529
|
-
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
|
|
530
|
-
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
|
|
531
|
-
|
|
532
|
-
float4 s = (float4)(0.f);
|
|
533
|
-
float smin = 0;
|
|
534
|
-
for (int l = 0; l < n; ++l) {
|
|
535
|
-
s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
|
|
536
|
-
s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
|
|
537
|
-
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
|
|
538
|
-
}
|
|
539
|
-
tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
|
|
540
|
-
|
|
541
|
-
}
|
|
542
|
-
|
|
543
|
-
// sum up partial sums and write back result
|
|
544
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
545
|
-
for (int s=16; s>0; s>>=1) {
|
|
546
|
-
if (tid < s) {
|
|
547
|
-
tmp[tid] += tmp[tid + s];
|
|
548
|
-
}
|
|
549
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
550
|
-
}
|
|
551
|
-
if (tid == 0) {
|
|
552
|
-
dst[row] = tmp[0];
|
|
553
|
-
}
|
|
554
|
-
}
|
|
555
|
-
|
|
556
|
-
__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
|
|
557
|
-
|
|
558
|
-
const uint16_t kmask1 = 0x3f3f;
|
|
559
|
-
const uint16_t kmask2 = 0x0f0f;
|
|
560
|
-
const uint16_t kmask3 = 0xc0c0;
|
|
561
|
-
|
|
562
|
-
const int row = get_group_id(0);
|
|
563
|
-
const int num_blocks_per_row = ncols / QK_K;
|
|
564
|
-
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
|
|
565
|
-
|
|
566
|
-
const int tid = get_local_id(0)/2; // 0...15
|
|
567
|
-
const int ix = get_local_id(0)%2;
|
|
568
|
-
|
|
569
|
-
const int il = tid/4; // 0...3
|
|
570
|
-
const int ir = tid - 4*il;// 0...3
|
|
571
|
-
const int n = 2;
|
|
572
|
-
|
|
573
|
-
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
|
|
574
|
-
const int in = il%2;
|
|
575
|
-
|
|
576
|
-
const int l0 = n*(2*ir + in);
|
|
577
|
-
const int q_offset = 32*im + l0;
|
|
578
|
-
const int y_offset = 64*im + l0;
|
|
579
|
-
|
|
580
|
-
const uint8_t hm1 = 1 << (2*im);
|
|
581
|
-
const uint8_t hm2 = hm1 << 4;
|
|
582
|
-
|
|
583
|
-
uint16_t aux[4];
|
|
584
|
-
const uint8_t * sc = (const uint8_t *)aux;
|
|
585
|
-
|
|
586
|
-
__global const struct block_q5_K * x = xx + ib0;
|
|
587
|
-
|
|
588
|
-
tmp[16 * ix + tid] = 0;
|
|
589
|
-
|
|
590
|
-
for (int i = ix; i < num_blocks_per_row; i += 2) {
|
|
591
|
-
|
|
592
|
-
__global const uint8_t * ql1 = x[i].qs + q_offset;
|
|
593
|
-
__global const uint8_t * ql2 = ql1 + 64;
|
|
594
|
-
__global const uint8_t * qh = x[i].qh + l0;
|
|
595
|
-
__global const float * y1 = yy + i*QK_K + y_offset;
|
|
596
|
-
__global const float * y2 = y1 + 128;
|
|
597
|
-
|
|
598
|
-
const float dall = vload_half(0, &x[i].d);
|
|
599
|
-
const float dmin = vload_half(0, &x[i].dmin);
|
|
600
|
-
|
|
601
|
-
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
|
|
602
|
-
aux[0] = a[im+0] & kmask1;
|
|
603
|
-
aux[1] = a[im+2] & kmask1;
|
|
604
|
-
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
|
|
605
|
-
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
|
|
606
|
-
|
|
607
|
-
float4 sum = (float4)(0.f);
|
|
608
|
-
float smin = 0;
|
|
609
|
-
for (int l = 0; l < n; ++l) {
|
|
610
|
-
sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
|
|
611
|
-
+ y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
|
|
612
|
-
sum.y += y1[l+32] * ((ql1[l+ 0] >> 4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
|
|
613
|
-
+ y1[l+48] * ((ql1[l+16] >> 4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
|
|
614
|
-
sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
|
|
615
|
-
+ y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
|
|
616
|
-
sum.w += y2[l+32] * ((ql2[l+ 0] >> 4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
|
|
617
|
-
+ y2[l+48] * ((ql2[l+16] >> 4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
|
|
618
|
-
smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
|
|
619
|
-
+ (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
|
|
620
|
-
}
|
|
621
|
-
tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
|
|
622
|
-
|
|
623
|
-
}
|
|
624
|
-
|
|
625
|
-
// sum up partial sums and write back result
|
|
626
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
627
|
-
for (int s=16; s>0; s>>=1) {
|
|
628
|
-
if (tid < s) {
|
|
629
|
-
tmp[tid] += tmp[tid + s];
|
|
630
|
-
}
|
|
631
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
632
|
-
}
|
|
633
|
-
if (tid == 0) {
|
|
634
|
-
dst[row] = tmp[0];
|
|
635
|
-
}
|
|
636
|
-
}
|
|
637
|
-
|
|
638
|
-
__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) {
|
|
639
|
-
|
|
640
|
-
const int row = get_group_id(0);
|
|
641
|
-
|
|
642
|
-
const int num_blocks_per_row = ncols / QK_K;
|
|
643
|
-
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
|
|
644
|
-
|
|
645
|
-
__global const struct block_q6_K * x = xx + ib0;
|
|
646
|
-
|
|
647
|
-
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
|
|
648
|
-
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
|
|
649
|
-
|
|
650
|
-
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
|
|
651
|
-
|
|
652
|
-
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
653
|
-
const int in = tid - step*im; // 0...15 or 0...7
|
|
654
|
-
|
|
655
|
-
\n#if K_QUANTS_PER_ITERATION == 1\n
|
|
656
|
-
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
|
|
657
|
-
const int is = 0;
|
|
658
|
-
|
|
659
|
-
\n#else\n
|
|
660
|
-
|
|
661
|
-
const int l0 = 4 * in; // 0, 4, 8, ..., 28
|
|
662
|
-
const int is = in / 4;
|
|
663
|
-
|
|
664
|
-
\n#endif\n
|
|
665
|
-
|
|
666
|
-
const int ql_offset = 64*im + l0;
|
|
667
|
-
const int qh_offset = 32*im + l0;
|
|
668
|
-
const int s_offset = 8*im + is;
|
|
669
|
-
const int y_offset = 128*im + l0;
|
|
670
|
-
|
|
671
|
-
tmp[16 * ix + tid] = 0; // partial sum for thread in warp
|
|
672
|
-
|
|
673
|
-
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
674
|
-
|
|
675
|
-
__global const float * y = yy + i * QK_K + y_offset;
|
|
676
|
-
__global const uint8_t * ql = x[i].ql + ql_offset;
|
|
677
|
-
__global const uint8_t * qh = x[i].qh + qh_offset;
|
|
678
|
-
__global const int8_t * s = x[i].scales + s_offset;
|
|
679
|
-
|
|
680
|
-
const float d = vload_half(0, &x[i].d);
|
|
681
|
-
|
|
682
|
-
\n#if K_QUANTS_PER_ITERATION == 1\n
|
|
683
|
-
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
|
|
684
|
-
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
|
|
685
|
-
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
|
|
686
|
-
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
|
|
687
|
-
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
|
|
688
|
-
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
|
|
689
|
-
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
|
|
690
|
-
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
|
|
691
|
-
tmp[16 * ix + tid] += sum;
|
|
692
|
-
\n#else\n
|
|
693
|
-
float sum = 0;
|
|
694
|
-
for (int l = 0; l < 4; ++l) {
|
|
695
|
-
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
|
|
696
|
-
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
|
|
697
|
-
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
|
|
698
|
-
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
|
|
699
|
-
}
|
|
700
|
-
tmp[16 * ix + tid] += sum;
|
|
701
|
-
\n#endif\n
|
|
702
|
-
|
|
703
|
-
}
|
|
704
|
-
|
|
705
|
-
// sum up partial sums and write back result
|
|
706
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
707
|
-
for (int s=16; s>0; s>>=1) {
|
|
708
|
-
if (tid < s) {
|
|
709
|
-
tmp[tid] += tmp[tid + s];
|
|
710
|
-
}
|
|
711
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
712
|
-
}
|
|
713
|
-
if (tid == 0) {
|
|
714
|
-
dst[row] = tmp[0];
|
|
715
|
-
}
|
|
716
|
-
}
|
|
717
|
-
);
|
|
718
|
-
|
|
719
|
-
|
|
720
|
-
std::string dequant_template = MULTILINE_QUOTE(
|
|
721
|
-
__kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
|
|
722
|
-
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
|
|
723
|
-
|
|
724
|
-
if (i >= get_global_size(0)) {
|
|
725
|
-
return;
|
|
726
|
-
}
|
|
727
|
-
|
|
728
|
-
const uint qk = QUANT_K;
|
|
729
|
-
const uint qr = QUANT_R;
|
|
730
|
-
|
|
731
|
-
const int ib = i/qk + get_global_offset(0); // block index
|
|
732
|
-
const int iqs = (i%qk)/qr; // quant index
|
|
733
|
-
const int iybs = i - i%qk; // y block start index
|
|
734
|
-
const int y_offset = qr == 1 ? 1 : qk/2;
|
|
735
|
-
|
|
736
|
-
// dequantize
|
|
737
|
-
float v0, v1;
|
|
738
|
-
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
|
|
739
|
-
y[iybs + iqs + 0] = v0;
|
|
740
|
-
y[iybs + iqs + y_offset] = v1;
|
|
741
|
-
}
|
|
742
|
-
);
|
|
743
|
-
|
|
744
|
-
std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
|
|
745
|
-
__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
|
|
746
|
-
const int local_size = get_local_size(0);
|
|
747
|
-
const int row = get_group_id(0);
|
|
748
|
-
const int tid = get_local_id(0);
|
|
749
|
-
|
|
750
|
-
const uint qk = QUANT_K;
|
|
751
|
-
const uint qr = QUANT_R;
|
|
752
|
-
|
|
753
|
-
const int col_step = local_size * 2;
|
|
754
|
-
const int y_offset = qr == 1 ? 1 : qk/2;
|
|
755
|
-
|
|
756
|
-
x += get_global_offset(0);
|
|
757
|
-
|
|
758
|
-
tmp[tid] = 0;
|
|
759
|
-
|
|
760
|
-
for (int col = tid*2; col < ncols; col += col_step) {
|
|
761
|
-
const int ib = (row*ncols + col)/qk; // block index
|
|
762
|
-
const int iqs = (col%qk)/qr; // quant index
|
|
763
|
-
const int iybs = col - col%qk; // y block start index
|
|
764
|
-
|
|
765
|
-
// dequantize
|
|
766
|
-
float v0, v1;
|
|
767
|
-
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
|
|
768
|
-
|
|
769
|
-
// matrix multiplication
|
|
770
|
-
tmp[tid] += v0 * y[iybs + iqs + 0];
|
|
771
|
-
tmp[tid] += v1 * y[iybs + iqs + y_offset];
|
|
772
|
-
}
|
|
773
|
-
|
|
774
|
-
// sum up partial sums and write back result
|
|
775
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
776
|
-
for (int s=local_size/2; s>0; s>>=1) {
|
|
777
|
-
if (tid < s) {
|
|
778
|
-
tmp[tid] += tmp[tid + s];
|
|
779
|
-
}
|
|
780
|
-
barrier(CLK_LOCAL_MEM_FENCE);
|
|
781
|
-
}
|
|
782
|
-
if (tid == 0) {
|
|
783
|
-
dst[row] = tmp[0];
|
|
784
|
-
}
|
|
785
|
-
}
|
|
786
|
-
|
|
787
|
-
);
|
|
788
|
-
|
|
789
|
-
|
|
790
|
-
std::string mul_template = MULTILINE_QUOTE(
|
|
791
|
-
__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
|
|
792
|
-
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
|
|
793
|
-
|
|
794
|
-
if (i >= get_global_size(0)) {
|
|
795
|
-
return;
|
|
796
|
-
}
|
|
797
|
-
|
|
798
|
-
dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
|
|
799
|
-
}
|
|
800
|
-
);
|
|
801
|
-
|
|
802
|
-
std::string add_template = MULTILINE_QUOTE(
|
|
803
|
-
__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) {
|
|
804
|
-
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
|
|
805
|
-
|
|
806
|
-
if (i >= get_global_size(0)) {
|
|
807
|
-
return;
|
|
808
|
-
}
|
|
809
|
-
|
|
810
|
-
dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky];
|
|
811
|
-
}
|
|
812
|
-
);
|
|
813
|
-
|
|
814
|
-
#define CL_CHECK(err) \
|
|
815
|
-
do { \
|
|
816
|
-
cl_int err_ = (err); \
|
|
817
|
-
if (err_ != CL_SUCCESS) { \
|
|
818
|
-
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
|
|
819
|
-
#err, err_, __FILE__, __LINE__); \
|
|
820
|
-
exit(1); \
|
|
821
|
-
} \
|
|
822
|
-
} while (0)
|
|
823
|
-
|
|
824
|
-
#define CLBLAST_CHECK(err) \
|
|
825
|
-
do { \
|
|
826
|
-
CLBlastStatusCode err_ = (err); \
|
|
827
|
-
if (err_ != CLBlastSuccess) { \
|
|
828
|
-
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
|
|
829
|
-
#err, err_, __FILE__, __LINE__); \
|
|
830
|
-
exit(1); \
|
|
831
|
-
} \
|
|
832
|
-
} while (0)
|
|
833
|
-
|
|
834
|
-
std::array<std::string, 5> dequant_str_keys = {
|
|
835
|
-
"KERNEL_NAME", "X_TYPE", "QUANT_K", "QUANT_R", "DEQUANT_FUNC"
|
|
836
|
-
};
|
|
837
|
-
|
|
838
|
-
std::array<std::string, 30> dequant_str_values = {
|
|
839
|
-
"dequantize_row_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
|
|
840
|
-
"dequantize_row_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
|
|
841
|
-
"dequantize_row_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
|
|
842
|
-
"dequantize_row_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
|
|
843
|
-
"dequantize_row_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
|
|
844
|
-
"convert_row_f16", "half", "1", "1", "convert_f16"
|
|
845
|
-
};
|
|
846
|
-
|
|
847
|
-
std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
|
|
848
|
-
"dequantize_mul_mat_vec_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
|
|
849
|
-
"dequantize_mul_mat_vec_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
|
|
850
|
-
"dequantize_mul_mat_vec_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
|
|
851
|
-
"dequantize_mul_mat_vec_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
|
|
852
|
-
"dequantize_mul_mat_vec_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
|
|
853
|
-
"convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
|
|
854
|
-
};
|
|
855
|
-
|
|
856
|
-
std::array<std::string, 2> mul_str_keys = {
|
|
857
|
-
"KERNEL_NAME", "TYPE"
|
|
858
|
-
};
|
|
859
|
-
std::array<std::string, 2> mul_str_values = {
|
|
860
|
-
"mul_f32", "float"
|
|
861
|
-
};
|
|
862
|
-
|
|
863
|
-
static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
|
|
864
|
-
size_t pos = 0;
|
|
865
|
-
while ((pos = s.find(from, pos)) != std::string::npos) {
|
|
866
|
-
s.replace(pos, from.length(), to);
|
|
867
|
-
pos += to.length();
|
|
868
|
-
}
|
|
869
|
-
return s;
|
|
870
|
-
}
|
|
871
|
-
|
|
872
|
-
static std::string generate_kernels() {
|
|
873
|
-
std::stringstream src;
|
|
874
|
-
src << program_source << '\n';
|
|
875
|
-
src << k_quants_source << '\n';
|
|
876
|
-
for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
|
|
877
|
-
std::string dequant_kernel = dequant_template;
|
|
878
|
-
std::string dmmv_kernel = dequant_mul_mat_vec_template;
|
|
879
|
-
for (size_t j = 0; j < dequant_str_keys.size(); j++) {
|
|
880
|
-
replace(dequant_kernel, dequant_str_keys[j], dequant_str_values[i + j]);
|
|
881
|
-
replace(dmmv_kernel, dequant_str_keys[j], dequant_mul_mat_vec_str_values[i + j]);
|
|
882
|
-
}
|
|
883
|
-
src << dequant_kernel << '\n';
|
|
884
|
-
src << dmmv_kernel << '\n';
|
|
885
|
-
}
|
|
886
|
-
for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
|
|
887
|
-
std::string mul_kernel = mul_template;
|
|
888
|
-
for (size_t j = 0; j < mul_str_keys.size(); j++) {
|
|
889
|
-
replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
|
|
890
|
-
}
|
|
891
|
-
src << mul_kernel << '\n';
|
|
892
|
-
}
|
|
893
|
-
src << add_template << '\n';
|
|
894
|
-
|
|
895
|
-
return src.str();
|
|
896
|
-
}
|
|
897
|
-
|
|
898
|
-
static cl_platform_id platform;
|
|
899
|
-
static cl_device_id device;
|
|
900
|
-
static cl_context context;
|
|
901
|
-
static cl_command_queue queue;
|
|
902
|
-
static cl_program program;
|
|
903
|
-
static cl_kernel convert_row_f16_cl;
|
|
904
|
-
static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
|
|
905
|
-
static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
|
|
906
|
-
static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
|
|
907
|
-
static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
|
|
908
|
-
static cl_kernel mul_f32_cl;
|
|
909
|
-
static cl_kernel add_f32_cl;
|
|
910
|
-
static bool fp16_support;
|
|
911
|
-
|
|
912
|
-
static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
|
|
913
|
-
cl_program p;
|
|
914
|
-
char *program_log;
|
|
915
|
-
size_t program_size;
|
|
916
|
-
size_t log_size;
|
|
917
|
-
int err;
|
|
918
|
-
|
|
919
|
-
program_size = strlen(program_buffer);
|
|
920
|
-
|
|
921
|
-
p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
|
|
922
|
-
if(err < 0) {
|
|
923
|
-
fprintf(stderr, "OpenCL error creating program");
|
|
924
|
-
exit(1);
|
|
925
|
-
}
|
|
926
|
-
|
|
927
|
-
std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
|
|
928
|
-
"-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 "
|
|
929
|
-
"-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION);
|
|
930
|
-
|
|
931
|
-
err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
|
|
932
|
-
if(err < 0) {
|
|
933
|
-
|
|
934
|
-
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
|
|
935
|
-
program_log = (char*) malloc(log_size + 1);
|
|
936
|
-
program_log[log_size] = '\0';
|
|
937
|
-
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
|
|
938
|
-
fprintf(stderr, "ggml_opencl: kernel compile error:\n\n%s\n", program_log);
|
|
939
|
-
free(program_log);
|
|
940
|
-
exit(1);
|
|
941
|
-
}
|
|
942
|
-
|
|
943
|
-
return p;
|
|
944
|
-
}
|
|
945
|
-
|
|
946
|
-
void ggml_cl_init(void) {
|
|
947
|
-
static bool initialized = false;
|
|
948
|
-
if (initialized) {
|
|
949
|
-
return;
|
|
950
|
-
}
|
|
951
|
-
initialized = true;
|
|
952
|
-
|
|
953
|
-
cl_int err;
|
|
954
|
-
|
|
955
|
-
struct cl_device;
|
|
956
|
-
struct cl_platform {
|
|
957
|
-
cl_platform_id id;
|
|
958
|
-
unsigned number;
|
|
959
|
-
char name[128];
|
|
960
|
-
char vendor[128];
|
|
961
|
-
struct cl_device * devices;
|
|
962
|
-
unsigned n_devices;
|
|
963
|
-
struct cl_device * default_device;
|
|
964
|
-
};
|
|
965
|
-
|
|
966
|
-
struct cl_device {
|
|
967
|
-
struct cl_platform * platform;
|
|
968
|
-
cl_device_id id;
|
|
969
|
-
unsigned number;
|
|
970
|
-
cl_device_type type;
|
|
971
|
-
char name[128];
|
|
972
|
-
};
|
|
973
|
-
|
|
974
|
-
enum { NPLAT = 16, NDEV = 16 };
|
|
975
|
-
|
|
976
|
-
struct cl_platform platforms[NPLAT];
|
|
977
|
-
unsigned n_platforms = 0;
|
|
978
|
-
struct cl_device devices[NDEV];
|
|
979
|
-
unsigned n_devices = 0;
|
|
980
|
-
struct cl_device * default_device = NULL;
|
|
981
|
-
|
|
982
|
-
platform = NULL;
|
|
983
|
-
device = NULL;
|
|
984
|
-
|
|
985
|
-
cl_platform_id platform_ids[NPLAT];
|
|
986
|
-
CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
|
|
987
|
-
|
|
988
|
-
for (unsigned i = 0; i < n_platforms; i++) {
|
|
989
|
-
struct cl_platform * p = &platforms[i];
|
|
990
|
-
p->number = i;
|
|
991
|
-
p->id = platform_ids[i];
|
|
992
|
-
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
|
|
993
|
-
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
|
|
994
|
-
|
|
995
|
-
cl_device_id device_ids[NDEV];
|
|
996
|
-
cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
|
|
997
|
-
if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
|
|
998
|
-
p->n_devices = 0;
|
|
999
|
-
} else {
|
|
1000
|
-
CL_CHECK(clGetDeviceIDsError);
|
|
1001
|
-
}
|
|
1002
|
-
p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
|
|
1003
|
-
p->default_device = NULL;
|
|
1004
|
-
|
|
1005
|
-
for (unsigned j = 0; j < p->n_devices; j++) {
|
|
1006
|
-
struct cl_device * d = &devices[n_devices];
|
|
1007
|
-
d->number = n_devices++;
|
|
1008
|
-
d->id = device_ids[j];
|
|
1009
|
-
d->platform = p;
|
|
1010
|
-
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
|
|
1011
|
-
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
|
|
1012
|
-
|
|
1013
|
-
if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
|
|
1014
|
-
p->default_device = d;
|
|
1015
|
-
}
|
|
1016
|
-
}
|
|
1017
|
-
|
|
1018
|
-
if (default_device == NULL && p->default_device != NULL) {
|
|
1019
|
-
default_device = p->default_device;
|
|
1020
|
-
}
|
|
1021
|
-
}
|
|
1022
|
-
|
|
1023
|
-
if (n_devices == 0) {
|
|
1024
|
-
fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
|
|
1025
|
-
exit(1);
|
|
1026
|
-
}
|
|
1027
|
-
|
|
1028
|
-
char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
|
|
1029
|
-
char * user_device_string = getenv("GGML_OPENCL_DEVICE");
|
|
1030
|
-
int user_platform_number = -1;
|
|
1031
|
-
int user_device_number = -1;
|
|
1032
|
-
|
|
1033
|
-
unsigned n;
|
|
1034
|
-
if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
|
|
1035
|
-
user_platform_number = (int)n;
|
|
1036
|
-
}
|
|
1037
|
-
if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
|
|
1038
|
-
user_device_number = (int)n;
|
|
1039
|
-
}
|
|
1040
|
-
if (user_platform_number != -1 && user_device_number != -1) {
|
|
1041
|
-
cl_platform* platform = &platforms[user_platform_number];
|
|
1042
|
-
if ((unsigned)user_device_number >= platform->n_devices) {
|
|
1043
|
-
fprintf(stderr, "ggml_opencl: invalid device number %d\n", user_device_number);
|
|
1044
|
-
exit(1);
|
|
1045
|
-
}
|
|
1046
|
-
default_device = &platform->devices[user_device_number];
|
|
1047
|
-
} else {
|
|
1048
|
-
|
|
1049
|
-
struct cl_device * selected_devices = devices;
|
|
1050
|
-
unsigned n_selected_devices = n_devices;
|
|
1051
|
-
|
|
1052
|
-
if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
|
|
1053
|
-
for (unsigned i = 0; i < n_platforms; i++) {
|
|
1054
|
-
struct cl_platform * p = &platforms[i];
|
|
1055
|
-
if (strstr(p->name, user_platform_string) != NULL ||
|
|
1056
|
-
strstr(p->vendor, user_platform_string) != NULL) {
|
|
1057
|
-
user_platform_number = (int)i;
|
|
1058
|
-
break;
|
|
1059
|
-
}
|
|
1060
|
-
}
|
|
1061
|
-
if (user_platform_number == -1) {
|
|
1062
|
-
fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
|
|
1063
|
-
exit(1);
|
|
1064
|
-
}
|
|
1065
|
-
}
|
|
1066
|
-
if (user_platform_number != -1) {
|
|
1067
|
-
struct cl_platform * p = &platforms[user_platform_number];
|
|
1068
|
-
selected_devices = p->devices;
|
|
1069
|
-
n_selected_devices = p->n_devices;
|
|
1070
|
-
default_device = p->default_device;
|
|
1071
|
-
if (n_selected_devices == 0) {
|
|
1072
|
-
fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
|
|
1073
|
-
exit(1);
|
|
1074
|
-
}
|
|
1075
|
-
}
|
|
1076
|
-
|
|
1077
|
-
if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
|
|
1078
|
-
for (unsigned i = 0; i < n_selected_devices; i++) {
|
|
1079
|
-
struct cl_device * d = &selected_devices[i];
|
|
1080
|
-
if (strstr(d->name, user_device_string) != NULL) {
|
|
1081
|
-
user_device_number = d->number;
|
|
1082
|
-
break;
|
|
1083
|
-
}
|
|
1084
|
-
}
|
|
1085
|
-
if (user_device_number == -1) {
|
|
1086
|
-
fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
|
|
1087
|
-
exit(1);
|
|
1088
|
-
}
|
|
1089
|
-
}
|
|
1090
|
-
if (user_device_number != -1) {
|
|
1091
|
-
selected_devices = &devices[user_device_number];
|
|
1092
|
-
n_selected_devices = 1;
|
|
1093
|
-
default_device = &selected_devices[0];
|
|
1094
|
-
}
|
|
1095
|
-
|
|
1096
|
-
GGML_ASSERT(n_selected_devices > 0);
|
|
1097
|
-
|
|
1098
|
-
if (default_device == NULL) {
|
|
1099
|
-
default_device = &selected_devices[0];
|
|
1100
|
-
}
|
|
1101
|
-
}
|
|
1102
|
-
|
|
1103
|
-
fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
|
|
1104
|
-
fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
|
|
1105
|
-
if (default_device->type != CL_DEVICE_TYPE_GPU) {
|
|
1106
|
-
fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
|
|
1107
|
-
}
|
|
1108
|
-
|
|
1109
|
-
platform = default_device->platform->id;
|
|
1110
|
-
device = default_device->id;
|
|
1111
|
-
|
|
1112
|
-
size_t ext_str_size;
|
|
1113
|
-
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
|
|
1114
|
-
char *ext_buffer = (char *)alloca(ext_str_size + 1);
|
|
1115
|
-
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
|
|
1116
|
-
ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
|
|
1117
|
-
// Disabled due to faulty outputs
|
|
1118
|
-
// Check if ext_buffer contains cl_khr_fp16
|
|
1119
|
-
fp16_support = false; // strstr(ext_buffer, "cl_khr_fp16") != NULL;
|
|
1120
|
-
// fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
|
|
1121
|
-
|
|
1122
|
-
cl_context_properties properties[] = {
|
|
1123
|
-
(intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
|
|
1124
|
-
};
|
|
1125
|
-
|
|
1126
|
-
CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
|
|
1127
|
-
|
|
1128
|
-
CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
|
|
1129
|
-
(err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
|
|
1130
|
-
(queue = clCreateCommandQueue(context, device, 0, &err), err)
|
|
1131
|
-
)));
|
|
1132
|
-
|
|
1133
|
-
const std::string kernel_src = generate_kernels();
|
|
1134
|
-
|
|
1135
|
-
program = build_program_from_source(context, device, kernel_src.c_str());
|
|
1136
|
-
|
|
1137
|
-
// FP16 to FP32 kernel
|
|
1138
|
-
CL_CHECK((convert_row_f16_cl = clCreateKernel(program, "convert_row_f16", &err), err));
|
|
1139
|
-
|
|
1140
|
-
// Dequantize kernels
|
|
1141
|
-
CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
|
|
1142
|
-
CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
|
|
1143
|
-
CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
|
|
1144
|
-
CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
|
|
1145
|
-
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
|
|
1146
|
-
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
|
|
1147
|
-
CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
|
|
1148
|
-
CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
|
|
1149
|
-
CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
|
|
1150
|
-
CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
|
|
1151
|
-
CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
|
|
1152
|
-
|
|
1153
|
-
// dequant mul mat kernel
|
|
1154
|
-
CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
|
|
1155
|
-
CL_CHECK((dequantize_mul_mat_vec_q4_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_1", &err), err));
|
|
1156
|
-
CL_CHECK((dequantize_mul_mat_vec_q5_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_0", &err), err));
|
|
1157
|
-
CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
|
|
1158
|
-
CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
|
|
1159
|
-
CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
|
|
1160
|
-
CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
|
|
1161
|
-
CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
|
|
1162
|
-
CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
|
|
1163
|
-
CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
|
|
1164
|
-
CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
|
|
1165
|
-
|
|
1166
|
-
// mul kernel
|
|
1167
|
-
CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
|
|
1168
|
-
|
|
1169
|
-
CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err));
|
|
1170
|
-
}
|
|
1171
|
-
|
|
1172
|
-
static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
|
|
1173
|
-
switch (type) {
|
|
1174
|
-
case GGML_TYPE_Q4_0:
|
|
1175
|
-
return &dequantize_row_q4_0_cl;
|
|
1176
|
-
case GGML_TYPE_Q4_1:
|
|
1177
|
-
return &dequantize_row_q4_1_cl;
|
|
1178
|
-
case GGML_TYPE_Q5_0:
|
|
1179
|
-
return &dequantize_row_q5_0_cl;
|
|
1180
|
-
case GGML_TYPE_Q5_1:
|
|
1181
|
-
return &dequantize_row_q5_1_cl;
|
|
1182
|
-
case GGML_TYPE_Q8_0:
|
|
1183
|
-
return &dequantize_row_q8_0_cl;
|
|
1184
|
-
case GGML_TYPE_Q2_K:
|
|
1185
|
-
return &dequantize_block_q2_k_cl;
|
|
1186
|
-
case GGML_TYPE_Q3_K:
|
|
1187
|
-
return &dequantize_block_q3_k_cl;
|
|
1188
|
-
case GGML_TYPE_Q4_K:
|
|
1189
|
-
return &dequantize_block_q4_k_cl;
|
|
1190
|
-
case GGML_TYPE_Q5_K:
|
|
1191
|
-
return &dequantize_block_q5_k_cl;
|
|
1192
|
-
case GGML_TYPE_Q6_K:
|
|
1193
|
-
return &dequantize_block_q6_k_cl;
|
|
1194
|
-
case GGML_TYPE_F16:
|
|
1195
|
-
return &convert_row_f16_cl;
|
|
1196
|
-
default:
|
|
1197
|
-
return nullptr;
|
|
1198
|
-
}
|
|
1199
|
-
}
|
|
1200
|
-
|
|
1201
|
-
static size_t ggml_cl_global_denom(ggml_type type) {
|
|
1202
|
-
switch (type) {
|
|
1203
|
-
case GGML_TYPE_Q4_0:
|
|
1204
|
-
case GGML_TYPE_Q4_1:
|
|
1205
|
-
case GGML_TYPE_Q5_0:
|
|
1206
|
-
case GGML_TYPE_Q5_1:
|
|
1207
|
-
case GGML_TYPE_Q8_0:
|
|
1208
|
-
return 1;
|
|
1209
|
-
case GGML_TYPE_Q2_K:
|
|
1210
|
-
case GGML_TYPE_Q3_K:
|
|
1211
|
-
return 4;
|
|
1212
|
-
case GGML_TYPE_Q4_K:
|
|
1213
|
-
return 8;
|
|
1214
|
-
case GGML_TYPE_Q5_K:
|
|
1215
|
-
case GGML_TYPE_Q6_K:
|
|
1216
|
-
return 4;
|
|
1217
|
-
case GGML_TYPE_F16:
|
|
1218
|
-
default:
|
|
1219
|
-
return 1;
|
|
1220
|
-
}
|
|
1221
|
-
}
|
|
1222
|
-
|
|
1223
|
-
static size_t ggml_cl_local_size(ggml_type type) {
|
|
1224
|
-
switch (type) {
|
|
1225
|
-
case GGML_TYPE_Q4_0:
|
|
1226
|
-
case GGML_TYPE_Q4_1:
|
|
1227
|
-
case GGML_TYPE_Q5_0:
|
|
1228
|
-
case GGML_TYPE_Q5_1:
|
|
1229
|
-
case GGML_TYPE_Q8_0:
|
|
1230
|
-
return 0;
|
|
1231
|
-
case GGML_TYPE_Q2_K:
|
|
1232
|
-
case GGML_TYPE_Q3_K:
|
|
1233
|
-
return 64;
|
|
1234
|
-
case GGML_TYPE_Q4_K:
|
|
1235
|
-
return 32;
|
|
1236
|
-
case GGML_TYPE_Q5_K:
|
|
1237
|
-
case GGML_TYPE_Q6_K:
|
|
1238
|
-
return 64;
|
|
1239
|
-
case GGML_TYPE_F16:
|
|
1240
|
-
default:
|
|
1241
|
-
return 0;
|
|
1242
|
-
}
|
|
1243
|
-
}
|
|
1244
|
-
|
|
1245
|
-
static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
|
|
1246
|
-
switch (type) {
|
|
1247
|
-
case GGML_TYPE_Q4_0:
|
|
1248
|
-
return &dequantize_mul_mat_vec_q4_0_cl;
|
|
1249
|
-
case GGML_TYPE_Q4_1:
|
|
1250
|
-
return &dequantize_mul_mat_vec_q4_1_cl;
|
|
1251
|
-
case GGML_TYPE_Q5_0:
|
|
1252
|
-
return &dequantize_mul_mat_vec_q5_0_cl;
|
|
1253
|
-
case GGML_TYPE_Q5_1:
|
|
1254
|
-
return &dequantize_mul_mat_vec_q5_1_cl;
|
|
1255
|
-
case GGML_TYPE_Q8_0:
|
|
1256
|
-
return &dequantize_mul_mat_vec_q8_0_cl;
|
|
1257
|
-
case GGML_TYPE_F16:
|
|
1258
|
-
return &convert_mul_mat_vec_f16_cl;
|
|
1259
|
-
case GGML_TYPE_Q2_K:
|
|
1260
|
-
return &dequantize_mul_mat_vec_q2_K_cl;
|
|
1261
|
-
case GGML_TYPE_Q3_K:
|
|
1262
|
-
return &dequantize_mul_mat_vec_q3_K_cl;
|
|
1263
|
-
case GGML_TYPE_Q4_K:
|
|
1264
|
-
return &dequantize_mul_mat_vec_q4_K_cl;
|
|
1265
|
-
case GGML_TYPE_Q5_K:
|
|
1266
|
-
return &dequantize_mul_mat_vec_q5_K_cl;
|
|
1267
|
-
case GGML_TYPE_Q6_K:
|
|
1268
|
-
return &dequantize_mul_mat_vec_q6_K_cl;
|
|
1269
|
-
default:
|
|
1270
|
-
return nullptr;
|
|
1271
|
-
}
|
|
1272
|
-
}
|
|
1273
|
-
|
|
1274
|
-
// buffer pool for cl
|
|
1275
|
-
#define MAX_CL_BUFFERS 256
|
|
1276
|
-
|
|
1277
|
-
struct scoped_spin_lock {
|
|
1278
|
-
std::atomic_flag& lock;
|
|
1279
|
-
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
|
|
1280
|
-
while (lock.test_and_set(std::memory_order_acquire)) {
|
|
1281
|
-
; // spin
|
|
1282
|
-
}
|
|
1283
|
-
}
|
|
1284
|
-
~scoped_spin_lock() {
|
|
1285
|
-
lock.clear(std::memory_order_release);
|
|
1286
|
-
}
|
|
1287
|
-
scoped_spin_lock(const scoped_spin_lock&) = delete;
|
|
1288
|
-
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
|
|
1289
|
-
};
|
|
1290
|
-
|
|
1291
|
-
struct cl_buffer {
|
|
1292
|
-
cl_mem mem;
|
|
1293
|
-
size_t size = 0;
|
|
1294
|
-
};
|
|
1295
|
-
|
|
1296
|
-
static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
|
|
1297
|
-
static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
|
|
1298
|
-
|
|
1299
|
-
static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
|
|
1300
|
-
scoped_spin_lock lock(g_cl_pool_lock);
|
|
1301
|
-
cl_int err;
|
|
1302
|
-
|
|
1303
|
-
int best_i = -1;
|
|
1304
|
-
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
|
|
1305
|
-
int worst_i = -1;
|
|
1306
|
-
size_t worst_size = 0; //largest unused buffer seen so far
|
|
1307
|
-
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
|
|
1308
|
-
cl_buffer &b = g_cl_buffer_pool[i];
|
|
1309
|
-
if (b.size > 0 && b.size >= size && b.size < best_size)
|
|
1310
|
-
{
|
|
1311
|
-
best_i = i;
|
|
1312
|
-
best_size = b.size;
|
|
1313
|
-
}
|
|
1314
|
-
if (b.size > 0 && b.size > worst_size)
|
|
1315
|
-
{
|
|
1316
|
-
worst_i = i;
|
|
1317
|
-
worst_size = b.size;
|
|
1318
|
-
}
|
|
1319
|
-
}
|
|
1320
|
-
if(best_i!=-1) //found the smallest buffer that fits our needs
|
|
1321
|
-
{
|
|
1322
|
-
cl_buffer& b = g_cl_buffer_pool[best_i];
|
|
1323
|
-
cl_mem mem = b.mem;
|
|
1324
|
-
*actual_size = b.size;
|
|
1325
|
-
b.size = 0;
|
|
1326
|
-
return mem;
|
|
1327
|
-
}
|
|
1328
|
-
if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
|
|
1329
|
-
{
|
|
1330
|
-
cl_buffer& b = g_cl_buffer_pool[worst_i];
|
|
1331
|
-
cl_mem mem = b.mem;
|
|
1332
|
-
b.size = 0;
|
|
1333
|
-
clReleaseMemObject(mem);
|
|
1334
|
-
}
|
|
1335
|
-
cl_mem mem;
|
|
1336
|
-
CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
|
|
1337
|
-
*actual_size = size;
|
|
1338
|
-
return mem;
|
|
1339
|
-
}
|
|
1340
|
-
|
|
1341
|
-
static void ggml_cl_pool_free(cl_mem mem, size_t size) {
|
|
1342
|
-
scoped_spin_lock lock(g_cl_pool_lock);
|
|
1343
|
-
|
|
1344
|
-
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
|
|
1345
|
-
cl_buffer& b = g_cl_buffer_pool[i];
|
|
1346
|
-
if (b.size == 0) {
|
|
1347
|
-
b.mem = mem;
|
|
1348
|
-
b.size = size;
|
|
1349
|
-
return;
|
|
1350
|
-
}
|
|
1351
|
-
}
|
|
1352
|
-
fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
|
|
1353
|
-
clReleaseMemObject(mem);
|
|
1354
|
-
}
|
|
1355
|
-
|
|
1356
|
-
void ggml_cl_free_data(const struct ggml_tensor* tensor) {
|
|
1357
|
-
if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
|
|
1358
|
-
return;
|
|
1359
|
-
}
|
|
1360
|
-
|
|
1361
|
-
cl_mem mem = (cl_mem)tensor->extra;
|
|
1362
|
-
clReleaseMemObject(mem);
|
|
1363
|
-
}
|
|
1364
|
-
|
|
1365
|
-
static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
|
|
1366
|
-
cl_int err;
|
|
1367
|
-
const uint64_t ne0 = src->ne[0];
|
|
1368
|
-
const uint64_t ne1 = src->ne[1];
|
|
1369
|
-
const uint64_t nb0 = src->nb[0];
|
|
1370
|
-
const uint64_t nb1 = src->nb[1];
|
|
1371
|
-
const uint64_t nb2 = src->nb[2];
|
|
1372
|
-
const uint64_t nb3 = src->nb[3];
|
|
1373
|
-
const enum ggml_type type = src->type;
|
|
1374
|
-
const size_t ts = ggml_type_size(type);
|
|
1375
|
-
const size_t bs = ggml_blck_size(type);
|
|
1376
|
-
const uint64_t row_size = ts*ne0/bs;
|
|
1377
|
-
|
|
1378
|
-
const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
|
|
1379
|
-
if (nb0 == ts && nb1 == row_size) {
|
|
1380
|
-
return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
|
|
1381
|
-
}
|
|
1382
|
-
if (nb0 == ts) {
|
|
1383
|
-
const size_t buffer_origin[3] = { offset, 0, 0 };
|
|
1384
|
-
const size_t host_origin[3] = { 0, 0, 0 };
|
|
1385
|
-
const size_t region[3] = { row_size, ne1, 1 };
|
|
1386
|
-
return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
|
|
1387
|
-
}
|
|
1388
|
-
std::vector<cl_event> events;
|
|
1389
|
-
if (ev && ne1>1) events.reserve(ne1-1);
|
|
1390
|
-
for (uint64_t i1 = 0; i1 < ne1; i1++) {
|
|
1391
|
-
// pretend the row is a matrix with cols=1
|
|
1392
|
-
const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
|
|
1393
|
-
const size_t host_origin[3] = { 0, 0, 0 };
|
|
1394
|
-
const size_t region[3] = { ts, ne0/bs, 1 };
|
|
1395
|
-
// if an event is requested, make the last write wait for all previous writes to complete
|
|
1396
|
-
if (ev && i1) {
|
|
1397
|
-
events.push_back(*ev);
|
|
1398
|
-
}
|
|
1399
|
-
cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
|
|
1400
|
-
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
|
|
1401
|
-
if (err != CL_SUCCESS) {
|
|
1402
|
-
for (auto event : events) {
|
|
1403
|
-
clReleaseEvent(event);
|
|
1404
|
-
}
|
|
1405
|
-
return err;
|
|
1406
|
-
}
|
|
1407
|
-
}
|
|
1408
|
-
for (auto event : events) {
|
|
1409
|
-
CL_CHECK(clReleaseEvent(event));
|
|
1410
|
-
}
|
|
1411
|
-
return CL_SUCCESS;
|
|
1412
|
-
}
|
|
1413
|
-
|
|
1414
|
-
static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
1415
|
-
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
|
|
1416
|
-
const int64_t ne00 = src0->ne[0];
|
|
1417
|
-
const int64_t ne01 = src0->ne[1];
|
|
1418
|
-
const int64_t ne02 = src0->ne[2];
|
|
1419
|
-
const int64_t ne03 = src0->ne[3];
|
|
1420
|
-
const int64_t ne10 = src1->ne[0];
|
|
1421
|
-
const int64_t ne11 = src1->ne[1];
|
|
1422
|
-
const int64_t ne12 = src1->ne[2];
|
|
1423
|
-
const int64_t ne13 = src1->ne[3];
|
|
1424
|
-
const int nb2 = dst->nb[2];
|
|
1425
|
-
const int nb3 = dst->nb[3];
|
|
1426
|
-
size_t x_size;
|
|
1427
|
-
size_t d_size;
|
|
1428
|
-
|
|
1429
|
-
cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
|
|
1430
|
-
cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
|
|
1431
|
-
cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
|
|
1432
|
-
|
|
1433
|
-
|
|
1434
|
-
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
|
1435
|
-
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
|
1436
|
-
cl_event ev;
|
|
1437
|
-
|
|
1438
|
-
// copy src0 to device
|
|
1439
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
|
|
1440
|
-
|
|
1441
|
-
const int64_t i13 = i03%ne13;
|
|
1442
|
-
const int64_t i12 = i02%ne12;
|
|
1443
|
-
const int i1 = i13*ne12*ne11 + i12*ne11;
|
|
1444
|
-
|
|
1445
|
-
cl_int x_offset = 0;
|
|
1446
|
-
cl_int y_offset = i1*ne10;
|
|
1447
|
-
cl_int d_offset = 0;
|
|
1448
|
-
|
|
1449
|
-
size_t global = ne00 * ne01;
|
|
1450
|
-
cl_int ky = ne10 * ne11;
|
|
1451
|
-
|
|
1452
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
|
|
1453
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
|
|
1454
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
|
|
1455
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
|
|
1456
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
|
|
1457
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
|
|
1458
|
-
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
|
|
1459
|
-
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
|
|
1460
|
-
|
|
1461
|
-
CL_CHECK(clReleaseEvent(ev));
|
|
1462
|
-
CL_CHECK(clFinish(queue));
|
|
1463
|
-
|
|
1464
|
-
// copy dst to host
|
|
1465
|
-
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
|
|
1466
|
-
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
|
|
1467
|
-
}
|
|
1468
|
-
}
|
|
1469
|
-
ggml_cl_pool_free(d_X, x_size);
|
|
1470
|
-
ggml_cl_pool_free(d_D, d_size);
|
|
1471
|
-
}
|
|
1472
|
-
|
|
1473
|
-
void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
|
|
1474
|
-
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
|
|
1475
|
-
ggml_cl_mul_f32(src0, src1, dst);
|
|
1476
|
-
}
|
|
1477
|
-
|
|
1478
|
-
static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
1479
|
-
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
|
|
1480
|
-
const int64_t ne00 = src0->ne[0];
|
|
1481
|
-
const int64_t ne01 = src0->ne[1];
|
|
1482
|
-
const int64_t ne02 = src0->ne[2];
|
|
1483
|
-
const int64_t ne03 = src0->ne[3];
|
|
1484
|
-
const int64_t ne10 = src1->ne[0];
|
|
1485
|
-
const int64_t ne11 = src1->ne[1];
|
|
1486
|
-
const int64_t ne12 = src1->ne[2];
|
|
1487
|
-
const int64_t ne13 = src1->ne[3];
|
|
1488
|
-
const int nb2 = dst->nb[2];
|
|
1489
|
-
const int nb3 = dst->nb[3];
|
|
1490
|
-
size_t x_size;
|
|
1491
|
-
size_t d_size;
|
|
1492
|
-
|
|
1493
|
-
cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
|
|
1494
|
-
cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
|
|
1495
|
-
cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
|
|
1496
|
-
|
|
1497
|
-
|
|
1498
|
-
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
|
1499
|
-
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
|
1500
|
-
cl_event ev;
|
|
1501
|
-
|
|
1502
|
-
// copy src0 to device
|
|
1503
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
|
|
1504
|
-
|
|
1505
|
-
const int64_t i13 = i03%ne13;
|
|
1506
|
-
const int64_t i12 = i02%ne12;
|
|
1507
|
-
const int i1 = i13*ne12*ne11 + i12*ne11;
|
|
1508
|
-
|
|
1509
|
-
cl_int x_offset = 0;
|
|
1510
|
-
cl_int y_offset = i1*ne10;
|
|
1511
|
-
cl_int d_offset = 0;
|
|
1512
|
-
|
|
1513
|
-
size_t global = ne00 * ne01;
|
|
1514
|
-
cl_int ky = ne10 * ne11;
|
|
1515
|
-
|
|
1516
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X));
|
|
1517
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset));
|
|
1518
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y));
|
|
1519
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset));
|
|
1520
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D));
|
|
1521
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset));
|
|
1522
|
-
CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky));
|
|
1523
|
-
CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
|
|
1524
|
-
|
|
1525
|
-
CL_CHECK(clReleaseEvent(ev));
|
|
1526
|
-
CL_CHECK(clFinish(queue));
|
|
1527
|
-
|
|
1528
|
-
// copy dst to host
|
|
1529
|
-
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
|
|
1530
|
-
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
|
|
1531
|
-
}
|
|
1532
|
-
}
|
|
1533
|
-
ggml_cl_pool_free(d_X, x_size);
|
|
1534
|
-
ggml_cl_pool_free(d_D, d_size);
|
|
1535
|
-
}
|
|
1536
|
-
|
|
1537
|
-
void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
|
|
1538
|
-
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
|
|
1539
|
-
ggml_cl_add_f32(src0, src1, dst);
|
|
1540
|
-
}
|
|
1541
|
-
|
|
1542
|
-
static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
1543
|
-
const int64_t ne00 = src0->ne[0];
|
|
1544
|
-
const int64_t ne01 = src0->ne[1];
|
|
1545
|
-
const int64_t ne02 = src0->ne[2];
|
|
1546
|
-
const int64_t ne03 = src0->ne[3];
|
|
1547
|
-
|
|
1548
|
-
const int64_t ne10 = src1->ne[0];
|
|
1549
|
-
const int64_t ne11 = src1->ne[1];
|
|
1550
|
-
const int64_t ne12 = src1->ne[2];
|
|
1551
|
-
const int64_t ne13 = src1->ne[3];
|
|
1552
|
-
|
|
1553
|
-
const int nb2 = dst->nb[2];
|
|
1554
|
-
const int nb3 = dst->nb[3];
|
|
1555
|
-
|
|
1556
|
-
const int64_t r2 = ne12 / ne02;
|
|
1557
|
-
const int64_t r3 = ne13 / ne03;
|
|
1558
|
-
|
|
1559
|
-
const float alpha = 1.0f;
|
|
1560
|
-
const float beta = 0.0f;
|
|
1561
|
-
const int x_ne = ne01 * ne00;
|
|
1562
|
-
const int y_ne = ne11 * ne10;
|
|
1563
|
-
const int d_ne = ne11 * ne01;
|
|
1564
|
-
|
|
1565
|
-
size_t x_size;
|
|
1566
|
-
size_t y_size;
|
|
1567
|
-
size_t d_size;
|
|
1568
|
-
cl_mem d_X;
|
|
1569
|
-
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
|
|
1570
|
-
d_X = (cl_mem) src0->extra;
|
|
1571
|
-
} else {
|
|
1572
|
-
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
|
|
1573
|
-
}
|
|
1574
|
-
cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
|
|
1575
|
-
cl_mem d_D = dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
|
|
1576
|
-
|
|
1577
|
-
size_t x_offset = 0;
|
|
1578
|
-
|
|
1579
|
-
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
|
1580
|
-
// TODO: copy src0 here when r3>1
|
|
1581
|
-
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
|
|
1582
|
-
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
|
1583
|
-
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
|
|
1584
|
-
x_offset = (i03 * ne02 + i02) * x_ne;
|
|
1585
|
-
} else {
|
|
1586
|
-
// copy src0 to device
|
|
1587
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
|
|
1588
|
-
}
|
|
1589
|
-
|
|
1590
|
-
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
|
|
1591
|
-
// copy src1 to device
|
|
1592
|
-
if (src1->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1593
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
|
|
1594
|
-
}
|
|
1595
|
-
|
|
1596
|
-
CL_CHECK(clFinish(queue));
|
|
1597
|
-
|
|
1598
|
-
// compute
|
|
1599
|
-
cl_event ev_sgemm;
|
|
1600
|
-
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
|
|
1601
|
-
clblast::Transpose::kYes, clblast::Transpose::kNo,
|
|
1602
|
-
ne01, ne11, ne10,
|
|
1603
|
-
alpha,
|
|
1604
|
-
d_X, x_offset, ne00,
|
|
1605
|
-
d_Y, 0, ne10,
|
|
1606
|
-
beta,
|
|
1607
|
-
d_D, 0, ne01,
|
|
1608
|
-
&queue, &ev_sgemm);
|
|
1609
|
-
|
|
1610
|
-
if (status != clblast::StatusCode::kSuccess) {
|
|
1611
|
-
GGML_ASSERT(false);
|
|
1612
|
-
}
|
|
1613
|
-
|
|
1614
|
-
// copy dst to host
|
|
1615
|
-
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1616
|
-
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
|
|
1617
|
-
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
|
|
1618
|
-
}
|
|
1619
|
-
}
|
|
1620
|
-
}
|
|
1621
|
-
}
|
|
1622
|
-
}
|
|
1623
|
-
|
|
1624
|
-
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
|
|
1625
|
-
ggml_cl_pool_free(d_X, x_size);
|
|
1626
|
-
}
|
|
1627
|
-
if (src1->backend != GGML_BACKEND_TYPE_GPU) {
|
|
1628
|
-
ggml_cl_pool_free(d_Y, y_size);
|
|
1629
|
-
}
|
|
1630
|
-
if (dst->backend != GGML_BACKEND_TYPE_GPU) {
|
|
1631
|
-
ggml_cl_pool_free(d_D, d_size);
|
|
1632
|
-
}
|
|
1633
|
-
}
|
|
1634
|
-
|
|
1635
|
-
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
|
|
1636
|
-
GGML_ASSERT(fp16_support);
|
|
1637
|
-
|
|
1638
|
-
const int64_t ne00 = src0->ne[0];
|
|
1639
|
-
const int64_t ne01 = src0->ne[1];
|
|
1640
|
-
const int64_t ne02 = src0->ne[2];
|
|
1641
|
-
const int64_t ne03 = src0->ne[3];
|
|
1642
|
-
|
|
1643
|
-
const int64_t ne10 = src1->ne[0];
|
|
1644
|
-
const int64_t ne11 = src1->ne[1];
|
|
1645
|
-
const int64_t ne12 = src1->ne[2];
|
|
1646
|
-
const int64_t ne13 = src1->ne[3];
|
|
1647
|
-
|
|
1648
|
-
const int nb10 = src1->nb[0];
|
|
1649
|
-
const int nb11 = src1->nb[1];
|
|
1650
|
-
const int nb12 = src1->nb[2];
|
|
1651
|
-
const int nb13 = src1->nb[3];
|
|
1652
|
-
|
|
1653
|
-
const int nb2 = dst->nb[2];
|
|
1654
|
-
const int nb3 = dst->nb[3];
|
|
1655
|
-
|
|
1656
|
-
const int64_t r2 = ne12 / ne02;
|
|
1657
|
-
const int64_t r3 = ne13 / ne03;
|
|
1658
|
-
|
|
1659
|
-
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
|
|
1660
|
-
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
|
|
1661
|
-
const int x_ne = ne01 * ne00;
|
|
1662
|
-
const int y_ne = ne11 * ne10;
|
|
1663
|
-
const int d_ne = ne11 * ne01;
|
|
1664
|
-
|
|
1665
|
-
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
|
|
1666
|
-
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
|
|
1667
|
-
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
|
|
1668
|
-
|
|
1669
|
-
size_t x_size;
|
|
1670
|
-
size_t y_size;
|
|
1671
|
-
size_t d_size;
|
|
1672
|
-
cl_mem d_X;
|
|
1673
|
-
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
|
|
1674
|
-
d_X = (cl_mem) src0->extra;
|
|
1675
|
-
} else {
|
|
1676
|
-
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
|
|
1677
|
-
}
|
|
1678
|
-
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
|
|
1679
|
-
cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
|
|
1680
|
-
|
|
1681
|
-
bool src1_cont_rows = nb10 == sizeof(float);
|
|
1682
|
-
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
|
|
1683
|
-
|
|
1684
|
-
size_t x_offset = 0;
|
|
1685
|
-
|
|
1686
|
-
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
|
1687
|
-
// TODO: copy src0 here when r3>1
|
|
1688
|
-
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
|
|
1689
|
-
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
|
1690
|
-
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
|
|
1691
|
-
x_offset = (i03 * ne02 + i02) * x_ne;
|
|
1692
|
-
} else {
|
|
1693
|
-
// copy src0 to device
|
|
1694
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
|
|
1695
|
-
}
|
|
1696
|
-
|
|
1697
|
-
// FIXME: convert on device
|
|
1698
|
-
|
|
1699
|
-
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
|
|
1700
|
-
// convert src1 to fp16
|
|
1701
|
-
// TODO: use multiple threads
|
|
1702
|
-
char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
|
|
1703
|
-
if (src1_cont_rows) {
|
|
1704
|
-
if (src1_cont_cols) {
|
|
1705
|
-
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
|
|
1706
|
-
}
|
|
1707
|
-
else {
|
|
1708
|
-
for (int64_t i11 = 0; i11 < ne11; i11++) {
|
|
1709
|
-
ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
|
|
1710
|
-
}
|
|
1711
|
-
}
|
|
1712
|
-
}
|
|
1713
|
-
else {
|
|
1714
|
-
for (int64_t i11 = 0; i11 < ne11; i11++) {
|
|
1715
|
-
for (int64_t i10 = 0; i10 < ne10; i10++) {
|
|
1716
|
-
// very slow due to no inlining
|
|
1717
|
-
tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
|
|
1718
|
-
}
|
|
1719
|
-
}
|
|
1720
|
-
}
|
|
1721
|
-
|
|
1722
|
-
// copy src1 to device
|
|
1723
|
-
CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
|
|
1724
|
-
|
|
1725
|
-
CL_CHECK(clFinish(queue));
|
|
1726
|
-
|
|
1727
|
-
// compute
|
|
1728
|
-
cl_event ev_sgemm;
|
|
1729
|
-
clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
|
|
1730
|
-
clblast::Transpose::kYes, clblast::Transpose::kNo,
|
|
1731
|
-
ne01, ne11, ne10,
|
|
1732
|
-
alpha,
|
|
1733
|
-
d_X, x_offset, ne00,
|
|
1734
|
-
d_Y, 0, ne10,
|
|
1735
|
-
beta,
|
|
1736
|
-
d_D, 0, ne01,
|
|
1737
|
-
&queue, &ev_sgemm);
|
|
1738
|
-
|
|
1739
|
-
if (status != clblast::StatusCode::kSuccess) {
|
|
1740
|
-
GGML_ASSERT(false);
|
|
1741
|
-
}
|
|
1742
|
-
|
|
1743
|
-
// copy dst to host, then convert to float
|
|
1744
|
-
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1745
|
-
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
|
|
1746
|
-
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
|
|
1747
|
-
ggml_fp16_to_fp32_row(tmp, d, d_ne);
|
|
1748
|
-
} else {
|
|
1749
|
-
// FIXME: convert dst to fp32 on device
|
|
1750
|
-
}
|
|
1751
|
-
}
|
|
1752
|
-
}
|
|
1753
|
-
}
|
|
1754
|
-
}
|
|
1755
|
-
|
|
1756
|
-
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
|
|
1757
|
-
ggml_cl_pool_free(d_X, x_size);
|
|
1758
|
-
}
|
|
1759
|
-
ggml_cl_pool_free(d_Y, y_size);
|
|
1760
|
-
ggml_cl_pool_free(d_D, d_size);
|
|
1761
|
-
}
|
|
1762
|
-
|
|
1763
|
-
static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
1764
|
-
const int64_t ne00 = src0->ne[0];
|
|
1765
|
-
const int64_t ne01 = src0->ne[1];
|
|
1766
|
-
const int64_t ne02 = src0->ne[2];
|
|
1767
|
-
const int64_t ne03 = src0->ne[3];
|
|
1768
|
-
|
|
1769
|
-
const int64_t ne10 = src1->ne[0];
|
|
1770
|
-
const int64_t ne11 = src1->ne[1];
|
|
1771
|
-
const int64_t ne12 = src1->ne[2];
|
|
1772
|
-
const int64_t ne13 = src1->ne[3];
|
|
1773
|
-
|
|
1774
|
-
const int nb2 = dst->nb[2];
|
|
1775
|
-
const int nb3 = dst->nb[3];
|
|
1776
|
-
const ggml_type type = src0->type;
|
|
1777
|
-
const bool mul_mat_vec = ne11 == 1 && ne00%2 == 0;
|
|
1778
|
-
|
|
1779
|
-
const int64_t r2 = ne12 / ne02;
|
|
1780
|
-
const int64_t r3 = ne13 / ne03;
|
|
1781
|
-
|
|
1782
|
-
const float alpha = 1.0f;
|
|
1783
|
-
const float beta = 0.0f;
|
|
1784
|
-
const int x_ne = ne01 * ne00;
|
|
1785
|
-
const int y_ne = ne11 * ne10;
|
|
1786
|
-
const int d_ne = ne11 * ne01;
|
|
1787
|
-
const int x_bps = x_ne / ggml_blck_size(type); // blocks per 2D slice
|
|
1788
|
-
const size_t q_sz = ggml_type_size(type) * x_bps;
|
|
1789
|
-
|
|
1790
|
-
size_t x_size;
|
|
1791
|
-
size_t y_size;
|
|
1792
|
-
size_t d_size;
|
|
1793
|
-
size_t q_size;
|
|
1794
|
-
cl_mem d_X;
|
|
1795
|
-
if (!mul_mat_vec) {
|
|
1796
|
-
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
|
|
1797
|
-
}
|
|
1798
|
-
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
|
|
1799
|
-
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
|
|
1800
|
-
cl_mem d_Q;
|
|
1801
|
-
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1802
|
-
d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
|
|
1803
|
-
}
|
|
1804
|
-
|
|
1805
|
-
cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
|
|
1806
|
-
cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
|
|
1807
|
-
GGML_ASSERT(to_fp32_cl != nullptr);
|
|
1808
|
-
|
|
1809
|
-
const size_t global_denom = ggml_cl_global_denom(type);
|
|
1810
|
-
const size_t local = mul_mat_vec ? CL_DMMV_LOCAL_SIZE : ggml_cl_local_size(type);
|
|
1811
|
-
|
|
1812
|
-
size_t ev_idx = 0;
|
|
1813
|
-
std::vector<cl_event> events;
|
|
1814
|
-
|
|
1815
|
-
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
|
1816
|
-
// TODO: copy and dequantize src0 here when r3>1
|
|
1817
|
-
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
|
|
1818
|
-
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
|
1819
|
-
// copy src0 to device if necessary
|
|
1820
|
-
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1821
|
-
events.emplace_back();
|
|
1822
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
|
|
1823
|
-
} else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
|
|
1824
|
-
d_Q = (cl_mem) src0->extra;
|
|
1825
|
-
} else {
|
|
1826
|
-
GGML_ASSERT(false);
|
|
1827
|
-
}
|
|
1828
|
-
|
|
1829
|
-
if (!mul_mat_vec) {
|
|
1830
|
-
// convert src0 to fp32 on device
|
|
1831
|
-
const size_t global = x_ne / global_denom;
|
|
1832
|
-
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
|
|
1833
|
-
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
|
|
1834
|
-
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
|
|
1835
|
-
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
|
|
1836
|
-
}
|
|
1837
|
-
|
|
1838
|
-
int64_t i12 = i02 * r2;
|
|
1839
|
-
int64_t e12 = i12 + r2;
|
|
1840
|
-
events.reserve(e12 - i12);
|
|
1841
|
-
for (; i12 < e12; i12++) {
|
|
1842
|
-
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
|
|
1843
|
-
// copy src1 to device
|
|
1844
|
-
events.emplace_back();
|
|
1845
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
|
|
1846
|
-
|
|
1847
|
-
// compute
|
|
1848
|
-
const size_t global = ne01 * local;
|
|
1849
|
-
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
|
|
1850
|
-
const cl_int ncols = ne00;
|
|
1851
|
-
events.emplace_back();
|
|
1852
|
-
CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
|
|
1853
|
-
CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
|
|
1854
|
-
CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
|
|
1855
|
-
CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
|
|
1856
|
-
CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
|
|
1857
|
-
CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
|
|
1858
|
-
} else { // CLBlast matrix matrix multiplication
|
|
1859
|
-
// copy src1 to device
|
|
1860
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
|
|
1861
|
-
|
|
1862
|
-
// wait for conversion
|
|
1863
|
-
CL_CHECK(clFinish(queue));
|
|
1864
|
-
|
|
1865
|
-
// compute
|
|
1866
|
-
events.emplace_back();
|
|
1867
|
-
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
|
|
1868
|
-
clblast::Transpose::kYes, clblast::Transpose::kNo,
|
|
1869
|
-
ne01, ne11, ne10,
|
|
1870
|
-
alpha,
|
|
1871
|
-
d_X, 0, ne00,
|
|
1872
|
-
d_Y, 0, ne10,
|
|
1873
|
-
beta,
|
|
1874
|
-
d_D, 0, ne01,
|
|
1875
|
-
&queue, events.data() + ev_idx++);
|
|
1876
|
-
|
|
1877
|
-
if (status != clblast::StatusCode::kSuccess) {
|
|
1878
|
-
GGML_ASSERT(false);
|
|
1879
|
-
}
|
|
1880
|
-
}
|
|
1881
|
-
|
|
1882
|
-
// copy dst to host
|
|
1883
|
-
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
|
|
1884
|
-
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
|
|
1885
|
-
for (auto *event : events) {
|
|
1886
|
-
clReleaseEvent(event);
|
|
1887
|
-
}
|
|
1888
|
-
|
|
1889
|
-
ev_idx = 0;
|
|
1890
|
-
events.clear();
|
|
1891
|
-
}
|
|
1892
|
-
}
|
|
1893
|
-
}
|
|
1894
|
-
}
|
|
1895
|
-
|
|
1896
|
-
if (!mul_mat_vec) {
|
|
1897
|
-
ggml_cl_pool_free(d_X, x_size);
|
|
1898
|
-
}
|
|
1899
|
-
ggml_cl_pool_free(d_Y, y_size);
|
|
1900
|
-
ggml_cl_pool_free(d_D, d_size);
|
|
1901
|
-
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
|
|
1902
|
-
ggml_cl_pool_free(d_Q, q_size);
|
|
1903
|
-
}
|
|
1904
|
-
}
|
|
1905
|
-
|
|
1906
|
-
|
|
1907
|
-
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst) {
|
|
1908
|
-
const int64_t ne10 = src1->ne[0];
|
|
1909
|
-
|
|
1910
|
-
const int64_t ne0 = dst->ne[0];
|
|
1911
|
-
const int64_t ne1 = dst->ne[1];
|
|
1912
|
-
|
|
1913
|
-
// TODO: find the optimal values for these
|
|
1914
|
-
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
|
|
1915
|
-
src1->type == GGML_TYPE_F32 &&
|
|
1916
|
-
dst->type == GGML_TYPE_F32 &&
|
|
1917
|
-
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
|
|
1918
|
-
return true;
|
|
1919
|
-
}
|
|
1920
|
-
|
|
1921
|
-
return false;
|
|
1922
|
-
}
|
|
1923
|
-
|
|
1924
|
-
static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
|
|
1925
|
-
// If device doesn't support FP16
|
|
1926
|
-
if (!fp16_support) {
|
|
1927
|
-
return false;
|
|
1928
|
-
}
|
|
1929
|
-
|
|
1930
|
-
size_t src0_sz = ggml_nbytes(src0);
|
|
1931
|
-
size_t src1_sz = ggml_nbytes(src1);
|
|
1932
|
-
|
|
1933
|
-
// mul_mat_q: src0 is converted to fp32 on device
|
|
1934
|
-
size_t mul_mat_q_transfer = src0_sz + src1_sz;
|
|
1935
|
-
|
|
1936
|
-
// mul_mat_f16: src1 is converted to fp16 on cpu
|
|
1937
|
-
size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
|
|
1938
|
-
|
|
1939
|
-
// choose the smaller one to transfer to the device
|
|
1940
|
-
// TODO: this is not always the best choice due to the overhead of converting to fp16
|
|
1941
|
-
return mul_mat_f16_transfer < mul_mat_q_transfer;
|
|
1942
|
-
}
|
|
1943
|
-
|
|
1944
|
-
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
|
|
1945
|
-
GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
|
|
1946
|
-
|
|
1947
|
-
if (src0->type == GGML_TYPE_F32) {
|
|
1948
|
-
ggml_cl_mul_mat_f32(src0, src1, dst);
|
|
1949
|
-
}
|
|
1950
|
-
else if (src0->type == GGML_TYPE_F16) {
|
|
1951
|
-
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
|
|
1952
|
-
ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
|
|
1953
|
-
}
|
|
1954
|
-
else {
|
|
1955
|
-
ggml_cl_mul_mat_q_f32(src0, src1, dst);
|
|
1956
|
-
}
|
|
1957
|
-
}
|
|
1958
|
-
else if (ggml_is_quantized(src0->type)) {
|
|
1959
|
-
ggml_cl_mul_mat_q_f32(src0, src1, dst);
|
|
1960
|
-
}
|
|
1961
|
-
else {
|
|
1962
|
-
GGML_ASSERT(false);
|
|
1963
|
-
}
|
|
1964
|
-
}
|
|
1965
|
-
|
|
1966
|
-
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
|
|
1967
|
-
if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
|
|
1968
|
-
return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
|
|
1969
|
-
}
|
|
1970
|
-
return 0;
|
|
1971
|
-
}
|
|
1972
|
-
|
|
1973
|
-
void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
|
|
1974
|
-
const int64_t ne0 = tensor->ne[0];
|
|
1975
|
-
const int64_t ne1 = tensor->ne[1];
|
|
1976
|
-
const int64_t ne2 = tensor->ne[2];
|
|
1977
|
-
const int64_t ne3 = tensor->ne[3];
|
|
1978
|
-
|
|
1979
|
-
const ggml_type type = tensor->type;
|
|
1980
|
-
const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
|
|
1981
|
-
const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
|
|
1982
|
-
|
|
1983
|
-
size_t q_size;
|
|
1984
|
-
cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
|
|
1985
|
-
|
|
1986
|
-
tensor->data = data;
|
|
1987
|
-
// copy tensor to device
|
|
1988
|
-
size_t offset = 0;
|
|
1989
|
-
for (int64_t i3 = 0; i3 < ne3; i3++) {
|
|
1990
|
-
for (int64_t i2 = 0; i2 < ne2; i2++) {
|
|
1991
|
-
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, offset, tensor, i3, i2, NULL));
|
|
1992
|
-
offset += s_sz;
|
|
1993
|
-
}
|
|
1994
|
-
}
|
|
1995
|
-
|
|
1996
|
-
CL_CHECK(clFinish(queue));
|
|
1997
|
-
|
|
1998
|
-
tensor->extra = dst;
|
|
1999
|
-
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
|
2000
|
-
}
|
|
2001
|
-
|
|
2002
|
-
// ggml-backend
|
|
2003
|
-
|
|
2004
|
-
// buffer
|
|
2005
|
-
|
|
2006
|
-
struct ggml_backend_opencl_buffer_context {
|
|
2007
|
-
~ggml_backend_opencl_buffer_context() {
|
|
2008
|
-
if (buffer) {
|
|
2009
|
-
clReleaseMemObject(buffer);
|
|
2010
|
-
}
|
|
2011
|
-
for (auto * sub_buffer : sub_buffers) {
|
|
2012
|
-
clReleaseMemObject(sub_buffer);
|
|
2013
|
-
}
|
|
2014
|
-
}
|
|
2015
|
-
|
|
2016
|
-
cl_mem buffer;
|
|
2017
|
-
std::vector<cl_mem> sub_buffers;
|
|
2018
|
-
};
|
|
2019
|
-
|
|
2020
|
-
static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
|
|
2021
|
-
|
|
2022
|
-
static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
|
|
2023
|
-
return "OpenCL";
|
|
2024
|
-
|
|
2025
|
-
GGML_UNUSED(buffer);
|
|
2026
|
-
}
|
|
2027
|
-
|
|
2028
|
-
static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
|
2029
|
-
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
|
|
2030
|
-
delete ctx;
|
|
2031
|
-
}
|
|
2032
|
-
|
|
2033
|
-
static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
|
|
2034
|
-
return cl_ptr_base;
|
|
2035
|
-
|
|
2036
|
-
GGML_UNUSED(buffer);
|
|
2037
|
-
}
|
|
2038
|
-
|
|
2039
|
-
static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
|
|
2040
|
-
if (tensor->view_src != NULL && tensor->view_offs == 0) {
|
|
2041
|
-
tensor->extra = tensor->view_src->extra;
|
|
2042
|
-
} else {
|
|
2043
|
-
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
|
|
2044
|
-
cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
|
|
2045
|
-
cl_int err;
|
|
2046
|
-
cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err);
|
|
2047
|
-
CL_CHECK(err);
|
|
2048
|
-
ctx->sub_buffers.push_back(sub_buffer);
|
|
2049
|
-
tensor->extra = sub_buffer;
|
|
2050
|
-
}
|
|
2051
|
-
tensor->backend = GGML_BACKEND_TYPE_GPU;
|
|
2052
|
-
}
|
|
2053
|
-
|
|
2054
|
-
static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
|
2055
|
-
cl_mem tensor_buffer = (cl_mem) tensor->extra;
|
|
2056
|
-
CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
|
|
2057
|
-
CL_CHECK(clFinish(queue));
|
|
2058
|
-
|
|
2059
|
-
GGML_UNUSED(buffer);
|
|
2060
|
-
}
|
|
2061
|
-
|
|
2062
|
-
static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
|
2063
|
-
cl_mem tensor_buffer = (cl_mem) tensor->extra;
|
|
2064
|
-
CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
|
|
2065
|
-
CL_CHECK(clFinish(queue));
|
|
2066
|
-
|
|
2067
|
-
GGML_UNUSED(buffer);
|
|
2068
|
-
}
|
|
2069
|
-
|
|
2070
|
-
static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
|
|
2071
|
-
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
|
|
2072
|
-
CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
|
|
2073
|
-
CL_CHECK(clFinish(queue));
|
|
2074
|
-
}
|
|
2075
|
-
|
|
2076
|
-
static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
|
|
2077
|
-
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
|
|
2078
|
-
for (auto * sub_buffer : ctx->sub_buffers) {
|
|
2079
|
-
clReleaseMemObject(sub_buffer);
|
|
2080
|
-
}
|
|
2081
|
-
ctx->sub_buffers.clear();
|
|
2082
|
-
}
|
|
2083
|
-
|
|
2084
|
-
static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
|
|
2085
|
-
/* .get_name = */ ggml_backend_opencl_buffer_get_name,
|
|
2086
|
-
/* .free_buffer = */ ggml_backend_opencl_buffer_free_buffer,
|
|
2087
|
-
/* .get_base = */ ggml_backend_opencl_buffer_get_base,
|
|
2088
|
-
/* .init_tensor = */ ggml_backend_opencl_buffer_init_tensor,
|
|
2089
|
-
/* .set_tensor = */ ggml_backend_opencl_buffer_set_tensor,
|
|
2090
|
-
/* .get_tensor = */ ggml_backend_opencl_buffer_get_tensor,
|
|
2091
|
-
/* .cpy_tensor = */ NULL,
|
|
2092
|
-
/* .clear = */ ggml_backend_opencl_buffer_clear,
|
|
2093
|
-
/* .reset = */ ggml_backend_opencl_buffer_reset,
|
|
2094
|
-
};
|
|
2095
|
-
|
|
2096
|
-
// buffer type
|
|
2097
|
-
|
|
2098
|
-
static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
|
|
2099
|
-
return "OpenCL";
|
|
2100
|
-
|
|
2101
|
-
GGML_UNUSED(buffer_type);
|
|
2102
|
-
}
|
|
2103
|
-
|
|
2104
|
-
static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
|
|
2105
|
-
ggml_cl_init();
|
|
2106
|
-
|
|
2107
|
-
cl_int err;
|
|
2108
|
-
cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
|
|
2109
|
-
if (err != CL_SUCCESS) {
|
|
2110
|
-
fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
|
|
2111
|
-
return nullptr;
|
|
2112
|
-
}
|
|
2113
|
-
|
|
2114
|
-
ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
|
|
2115
|
-
|
|
2116
|
-
return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
|
|
2117
|
-
}
|
|
2118
|
-
|
|
2119
|
-
static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
|
|
2120
|
-
// FIXME: not thread safe, device may not be initialized yet
|
|
2121
|
-
static cl_uint alignment = -1;
|
|
2122
|
-
if (alignment == (cl_uint)-1) {
|
|
2123
|
-
ggml_cl_init();
|
|
2124
|
-
clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
|
|
2125
|
-
alignment /= 8; // bits to bytes
|
|
2126
|
-
}
|
|
2127
|
-
return alignment;
|
|
2128
|
-
|
|
2129
|
-
GGML_UNUSED(buffer_type);
|
|
2130
|
-
}
|
|
2131
|
-
|
|
2132
|
-
static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
|
|
2133
|
-
static size_t max_size = -1;
|
|
2134
|
-
if (max_size == (size_t)-1) {
|
|
2135
|
-
ggml_cl_init();
|
|
2136
|
-
clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_size, NULL);
|
|
2137
|
-
}
|
|
2138
|
-
return max_size;
|
|
2139
|
-
}
|
|
2140
|
-
|
|
2141
|
-
static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
|
|
2142
|
-
//return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
|
|
2143
|
-
return ggml_backend_is_cpu(backend);
|
|
2144
|
-
|
|
2145
|
-
GGML_UNUSED(buffer_type);
|
|
2146
|
-
}
|
|
2147
|
-
|
|
2148
|
-
static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
|
|
2149
|
-
/* .get_name = */ ggml_backend_opencl_buffer_type_name,
|
|
2150
|
-
/* .alloc_buffer = */ ggml_backend_opencl_buffer_type_alloc_buffer,
|
|
2151
|
-
/* .get_alignment = */ ggml_backend_opencl_buffer_type_get_alignment,
|
|
2152
|
-
/* .get_max_size = */ ggml_backend_opencl_buffer_type_get_max_size,
|
|
2153
|
-
/* .get_alloc_size = */ NULL,
|
|
2154
|
-
/* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
|
|
2155
|
-
/* .is_host = */ NULL,
|
|
2156
|
-
};
|
|
2157
|
-
|
|
2158
|
-
|
|
2159
|
-
ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
|
|
2160
|
-
static ggml_backend_buffer_type buffer_type = {
|
|
2161
|
-
/* .iface = */ ggml_backend_opencl_buffer_type_interface,
|
|
2162
|
-
/* .context = */ nullptr,
|
|
2163
|
-
};
|
|
2164
|
-
|
|
2165
|
-
return &buffer_type;
|
|
2166
|
-
}
|
|
2167
|
-
|
|
2168
|
-
#if 0
|
|
2169
|
-
// host buffer type
|
|
2170
|
-
|
|
2171
|
-
static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
|
2172
|
-
return "CL_Host";
|
|
2173
|
-
|
|
2174
|
-
GGML_UNUSED(buft);
|
|
2175
|
-
}
|
|
2176
|
-
|
|
2177
|
-
static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
|
|
2178
|
-
return "CL_Host";
|
|
2179
|
-
|
|
2180
|
-
GGML_UNUSED(buffer);
|
|
2181
|
-
}
|
|
2182
|
-
|
|
2183
|
-
static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
|
2184
|
-
ggml_cl_host_free(buffer->context);
|
|
2185
|
-
}
|
|
2186
|
-
|
|
2187
|
-
static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
|
2188
|
-
void * ptr = ggml_cl_host_malloc(size);
|
|
2189
|
-
|
|
2190
|
-
if (ptr == nullptr) {
|
|
2191
|
-
// fallback to cpu buffer
|
|
2192
|
-
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
|
|
2193
|
-
}
|
|
2194
|
-
|
|
2195
|
-
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
|
|
2196
|
-
buffer->buft = buft;
|
|
2197
|
-
buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
|
|
2198
|
-
buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
|
|
2199
|
-
|
|
2200
|
-
return buffer;
|
|
2201
|
-
}
|
|
2202
|
-
|
|
2203
|
-
ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
|
|
2204
|
-
static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
|
|
2205
|
-
/* .iface = */ {
|
|
2206
|
-
/* .get_name = */ ggml_backend_opencl_host_buffer_type_name,
|
|
2207
|
-
/* .alloc_buffer = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
|
|
2208
|
-
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
|
|
2209
|
-
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
|
|
2210
|
-
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
|
|
2211
|
-
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
|
|
2212
|
-
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
|
|
2213
|
-
},
|
|
2214
|
-
/* .context = */ nullptr,
|
|
2215
|
-
};
|
|
2216
|
-
|
|
2217
|
-
return &ggml_backend_opencl_buffer_type_host;
|
|
2218
|
-
}
|
|
2219
|
-
|
|
2220
|
-
// backend
|
|
2221
|
-
|
|
2222
|
-
static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
|
|
2223
|
-
return "OpenCL";
|
|
2224
|
-
|
|
2225
|
-
GGML_UNUSED(backend);
|
|
2226
|
-
}
|
|
2227
|
-
|
|
2228
|
-
static void ggml_backend_opencl_free(ggml_backend_t backend) {
|
|
2229
|
-
GGML_UNUSED(backend);
|
|
2230
|
-
}
|
|
2231
|
-
|
|
2232
|
-
static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
|
|
2233
|
-
return ggml_backend_opencl_buffer_type();
|
|
2234
|
-
|
|
2235
|
-
GGML_UNUSED(backend);
|
|
2236
|
-
}
|
|
2237
|
-
|
|
2238
|
-
static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
|
|
2239
|
-
for (int i = 0; i < graph->n_nodes; ++i) {
|
|
2240
|
-
ggml_tensor * node = graph->nodes[i];
|
|
2241
|
-
|
|
2242
|
-
if (ggml_is_empty(node)) {
|
|
2243
|
-
continue;
|
|
2244
|
-
}
|
|
2245
|
-
|
|
2246
|
-
switch (node->op) {
|
|
2247
|
-
case GGML_OP_MUL_MAT:
|
|
2248
|
-
ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
|
|
2249
|
-
break;
|
|
2250
|
-
case GGML_OP_MUL:
|
|
2251
|
-
ggml_cl_mul(node->src[0], node->src[1], node);
|
|
2252
|
-
break;
|
|
2253
|
-
default:
|
|
2254
|
-
GGML_ASSERT(false);
|
|
2255
|
-
}
|
|
2256
|
-
}
|
|
2257
|
-
|
|
2258
|
-
return GGML_STATUS_SUCCESS;
|
|
2259
|
-
|
|
2260
|
-
GGML_UNUSED(backend);
|
|
2261
|
-
}
|
|
2262
|
-
|
|
2263
|
-
static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
|
|
2264
|
-
switch (op->op) {
|
|
2265
|
-
case GGML_OP_MUL_MAT:
|
|
2266
|
-
return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
|
|
2267
|
-
case GGML_OP_MUL:
|
|
2268
|
-
// return ggml_can_repeat_rows(op->src[1], op->src[0]);
|
|
2269
|
-
return true;
|
|
2270
|
-
default:
|
|
2271
|
-
return false;
|
|
2272
|
-
}
|
|
2273
|
-
|
|
2274
|
-
GGML_UNUSED(backend);
|
|
2275
|
-
}
|
|
2276
|
-
|
|
2277
|
-
static ggml_backend_i opencl_backend_i = {
|
|
2278
|
-
/* .get_name = */ ggml_backend_opencl_name,
|
|
2279
|
-
/* .free = */ ggml_backend_opencl_free,
|
|
2280
|
-
/* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
|
|
2281
|
-
/* .set_tensor_async = */ NULL,
|
|
2282
|
-
/* .get_tensor_async = */ NULL,
|
|
2283
|
-
/* .cpy_tensor_from_async = */ NULL,
|
|
2284
|
-
/* .cpy_tensor_to_async = */ NULL,
|
|
2285
|
-
/* .synchronize = */ NULL,
|
|
2286
|
-
/* .graph_plan_create = */ NULL,
|
|
2287
|
-
/* .graph_plan_free = */ NULL,
|
|
2288
|
-
/* .graph_plan_compute = */ NULL,
|
|
2289
|
-
/* .graph_compute = */ ggml_backend_opencl_graph_compute,
|
|
2290
|
-
/* .supports_op = */ ggml_backend_opencl_supports_op,
|
|
2291
|
-
};
|
|
2292
|
-
|
|
2293
|
-
ggml_backend_t ggml_backend_opencl_init() {
|
|
2294
|
-
ggml_backend_t backend = new ggml_backend {
|
|
2295
|
-
/* .interface = */ opencl_backend_i,
|
|
2296
|
-
/* .context = */ nullptr
|
|
2297
|
-
};
|
|
2298
|
-
|
|
2299
|
-
return backend;
|
|
2300
|
-
}
|
|
2301
|
-
|
|
2302
|
-
bool ggml_backend_is_opencl(ggml_backend_t backend) {
|
|
2303
|
-
return backend && backend->iface.get_name == ggml_backend_opencl_name;
|
|
2304
|
-
}
|
|
2305
|
-
#endif
|