char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Jeff Bolzvulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (#21572)1f30ac0ce

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.6 KiB124 linesraw
1#version 450
2
3#extension GL_EXT_shader_16bit_storage : require
4#extension GL_EXT_control_flow_attributes : require
5#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
6
7#include "types.glsl"
8
9layout (push_constant) uniform parameter
10{
11    BDA_STORAGE_T dst_addr;
12    uint32_t nb10;
13    uint32_t nb11;
14    uint32_t nb12;
15    uint32_t nb13;
16    uint32_t s0;
17    uint32_t s1;
18    uint32_t s2;
19    uint32_t p0;
20    uint32_t p1;
21    uint32_t p2;
22    uint32_t d0;
23    uint32_t d1;
24    uint32_t d2;
25    uint32_t IW;
26    uint32_t IH;
27    uint32_t ID;
28    uint32_t IC;
29    uint32_t KW;
30    uint32_t OH;
31    uint32_t KD_KH_KW;
32    uint32_t KH_KW;
33    uint32_t IC_KD_KH_KW;
34    uint32_t N_OD_OH;
35    uint32_t OD_OH;
36    uint32_t OD_OH_OW_IC_KD_KH_KW;
37    uint32_t OH_OW_IC_KD_KH_KW;
38    uint32_t OW_IC_KD_KH_KW;
39    uint32_t misalign_offsets;
40} p;
41
42uint get_aoffset() { return p.misalign_offsets >> 16; }
43uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
44
45layout(constant_id = 0) const uint BLOCK_SIZE = 32;
46
47layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
48
49layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
50layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
51
52#if BDA
53layout (buffer_reference) buffer D_ptr {D_TYPE d;};
54#endif
55
56void main() {
57    const uint32_t i = gl_GlobalInvocationID.x;
58
59    uint32_t nb10 = p.nb10;
60    uint32_t nb11 = p.nb11;
61    uint32_t nb12 = p.nb12;
62    uint32_t nb13 = p.nb13;
63    uint32_t s0 = p.s0;
64    uint32_t s1 = p.s1;
65    uint32_t s2 = p.s2;
66    uint32_t p0 = p.p0;
67    uint32_t p1 = p.p1;
68    uint32_t p2 = p.p2;
69    uint32_t d0 = p.d0;
70    uint32_t d1 = p.d1;
71    uint32_t d2 = p.d2;
72    uint32_t IW = p.IW;
73    uint32_t IH = p.IH;
74    uint32_t ID = p.ID;
75    uint32_t IC = p.IC;
76    uint32_t KW = p.KW;
77    uint32_t OH = p.OH;
78    uint32_t KD_KH_KW = p.KD_KH_KW;
79    uint32_t KH_KW = p.KH_KW;
80    uint32_t IC_KD_KH_KW = p.IC_KD_KH_KW;
81    uint32_t N_OD_OH = p.N_OD_OH;
82    uint32_t OD_OH = p.OD_OH;
83    uint32_t OD_OH_OW_IC_KD_KH_KW = p.OD_OH_OW_IC_KD_KH_KW;
84    uint32_t OH_OW_IC_KD_KH_KW = p.OH_OW_IC_KD_KH_KW;
85    uint32_t OW_IC_KD_KH_KW = p.OW_IC_KD_KH_KW;
86
87    if (i >= IC_KD_KH_KW) {
88        return;
89    }
90
91    const uint32_t iic = i / KD_KH_KW;
92    const uint32_t ikd = (i - iic * KD_KH_KW) / KH_KW;
93    const uint32_t ikh = (i - iic * KD_KH_KW - ikd * KH_KW) / KW;
94    const uint32_t ikw = i % KW;
95
96    const uint32_t iow = gl_GlobalInvocationID.y;
97    for (uint32_t iz = gl_GlobalInvocationID.z; iz < N_OD_OH; iz += gl_NumWorkGroups.z) {
98        const uint32_t in_ = iz / OD_OH;
99        const uint32_t iod = (iz - in_*OD_OH) / OH;
100        const uint32_t ioh = iz % OH;
101
102        const uint32_t iiw = iow * s0 + ikw * d0 - p0;
103        const uint32_t iih = ioh * s1 + ikh * d1 - p1;
104        const uint32_t iid = iod * s2 + ikd * d2 - p2;
105
106        const BDA_OFFSET_T offset_dst = BDA_OFFSET_T(in_)*OD_OH_OW_IC_KD_KH_KW + BDA_OFFSET_T(iod)*OH_OW_IC_KD_KH_KW + BDA_OFFSET_T(ioh)*OW_IC_KD_KH_KW + BDA_OFFSET_T(iow)*IC_KD_KH_KW + iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw;
107
108        const uint32_t offset_src = (in_*IC + iic)*nb13 + iid*nb12 + iih*nb11 + iiw*nb10;
109#if BDA
110        D_ptr dst_addr = D_ptr(p.dst_addr + D_SIZE * offset_dst);
111        if (iih >= IH || iiw >= IW || iid >= ID) {
112            dst_addr.d = D_TYPE(0.0f);
113        } else {
114            dst_addr.d = D_TYPE(data_a[offset_src + get_aoffset()]);
115        }
116#else
117        if (iih >= IH || iiw >= IW || iid >= ID) {
118            data_d[offset_dst + get_doffset()] = D_TYPE(0.0f);
119        } else {
120            data_d[offset_dst + get_doffset()] = D_TYPE(data_a[offset_src + get_aoffset()]);
121        }
122#endif
123    }
124}