@@ -16,48 +16,48 @@ typedef uchar uint8_t;
16
16
typedef int int32_t ;
17
17
typedef uint uint32_t ;
18
18
19
- constant uint QK4_0 = 32 ;
19
+ // constant uint QK4_0 = 32;
20
20
struct block_q4_0
21
21
{
22
22
float d ;
23
- uint8_t qs [QK4_0 / 2 ];
23
+ uint8_t qs [16 ]; // QK4_0 / 2
24
24
};
25
25
26
- constant uint QK4_1 = 32 ;
26
+ // constant uint QK4_1 = 32;
27
27
struct block_q4_1
28
28
{
29
29
float d ;
30
30
float m ;
31
- uint8_t qs [QK4_1 / 2 ];
31
+ uint8_t qs [16 ]; // QK4_1 / 2
32
32
};
33
33
34
- constant uint QK5_0 = 32 ;
34
+ // constant uint QK5_0 = 32;
35
35
struct __attribute__ ((packed )) block_q5_0
36
36
{
37
37
half d ;
38
38
uint32_t qh ;
39
- uint8_t qs [QK5_0 / 2 ];
39
+ uint8_t qs [16 ]; // QK5_0 / 2
40
40
};
41
41
42
- constant uint QK5_1 = 32 ;
42
+ // constant uint QK5_1 = 32;
43
43
struct block_q5_1
44
44
{
45
45
half d ;
46
46
half m ;
47
47
uint32_t qh ;
48
- uint8_t qs [QK5_1 / 2 ];
48
+ uint8_t qs [16 ]; // QK5_1 / 2
49
49
};
50
50
51
- constant uint QK8_0 = 32 ;
51
+ // constant uint QK8_0 = 32;
52
52
struct block_q8_0
53
53
{
54
54
float d ;
55
- uint8_t qs [QK8_0 ];
55
+ uint8_t qs [16 ]; // QK8_0 / 2
56
56
};
57
57
58
58
59
59
__kernel void dequantize_row_q4_0 (__global struct block_q4_0 * x , __global float * y ) {
60
- constant uint qk = QK4_0 ;
60
+ constant uint qk = 32 ; // QK4_0;
61
61
62
62
const uint i = get_global_id (0 ) / qk ;
63
63
const uint j = get_local_id (0 );
@@ -72,7 +72,7 @@ __kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float*
72
72
}
73
73
74
74
__kernel void dequantize_row_q4_1 (__global struct block_q4_1 * x , __global float * y ) {
75
- constant uint qk = QK4_1 ;
75
+ constant uint qk = 32 ; // QK4_1;
76
76
77
77
const uint i = get_global_id (0 ) / qk ;
78
78
const uint j = get_local_id (0 );
@@ -88,7 +88,7 @@ __kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float*
88
88
}
89
89
90
90
__kernel void dequantize_row_q5_0 (__global struct block_q5_0 * x , __global float * y ) {
91
- constant uint qk = QK5_0 ;
91
+ constant uint qk = 32 ; // QK5_0;
92
92
93
93
const uint i = get_global_id (0 ) / qk ;
94
94
const uint j = get_local_id (0 );
@@ -108,7 +108,7 @@ __kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float*
108
108
}
109
109
110
110
__kernel void dequantize_row_q5_1 (__global struct block_q5_1 * x , __global float * y ) {
111
- constant uint qk = QK5_1 ;
111
+ constant uint qk = 32 ; // QK5_1;
112
112
113
113
const uint i = get_global_id (0 ) / qk ;
114
114
const uint j = get_local_id (0 );
@@ -129,7 +129,7 @@ __kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float*
129
129
}
130
130
131
131
__kernel void dequantize_row_q8_0 (__global struct block_q8_0 * x , __global float * y ) {
132
- constant uint qk = QK8_0 ;
132
+ constant uint qk = 32 ; // QK8_0;
133
133
const uint i = get_global_id (0 ) / qk ;
134
134
const uint j = get_local_id (0 );
135
135
0 commit comments