mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-12-25 10:54:36 +00:00
7c7836d9d4
* Refactor shaders, extract GLSL code from ggml_vk_generate_shaders.py into vulkan-shaders directory * Improve debug log code * Add memory debug output option * Fix flake8 * Fix unnecessary high llama-3 VRAM use
61 lines
2.1 KiB
Plaintext
61 lines
2.1 KiB
Plaintext
#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
|
|
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
|
|
#endif
|
|
|
|
#if defined(DATA_A_F32)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_F16)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_Q4_0)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
const float d = float(data_a[a_offset + ib].d);
|
|
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
|
|
return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_Q4_1)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
const float d = float(data_a[a_offset + ib].d);
|
|
const float m = float(data_a[a_offset + ib].m);
|
|
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
|
|
return vec2(vui & 0xF, vui >> 4) * d + m;
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_Q5_0)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
const float d = float(data_a[a_offset + ib].d);
|
|
const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
|
|
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
|
|
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
|
|
return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_Q5_1)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
const float d = float(data_a[a_offset + ib].d);
|
|
const float m = float(data_a[a_offset + ib].m);
|
|
const uint uint_qh = data_a[a_offset + ib].qh;
|
|
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
|
|
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
|
|
return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
|
|
}
|
|
#endif
|
|
|
|
#if defined(DATA_A_Q8_0)
|
|
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
|
|
const float d = float(data_a[a_offset + ib].d);
|
|
return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
|
|
}
|
|
#endif
|