Spaces:
Build error
Build error
File size: 10,659 Bytes
5bd179e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 |
#ifndef _matrix_cuh
#define _matrix_cuh
#include <cuda_runtime.h>
#include <cuda_fp16.h>
//#include "cuda_buffers.cuh"
class MatrixView_half
{
public:
const half* data;
const int height;
const int width;
__device__ __forceinline__ MatrixView_half(const half* data, const int height, const int width)
: data(data), height(height), width(width)
{ }
__device__ __forceinline__ half item(int row, int column) const { return data[row * width + column]; }
__device__ __forceinline__ half2 item_half2(int row, int column) const { return ((half2*)data)[(row * width + column) / 2]; }
__device__ __forceinline__ half2 item_half2half2(int row, int column) const { return __half2half2(data[row * width + column]); }
__device__ __forceinline__ const half* item_ptr(int row, int column) const { return &data[row * width + column]; }
};
class MatrixView_half_rw
{
public:
half* data;
const int height;
const int width;
__device__ __forceinline__ MatrixView_half_rw(half* data, const int height, const int width)
: data(data), height(height), width(width)
{ }
__device__ __forceinline__ half item(int row, int column) const { return data[row * width + column]; }
__device__ __forceinline__ half2 item_half2(int row, int column) const { return ((half2*)data)[(row * width + column) / 2]; }
__device__ __forceinline__ half* item_ptr(int row, int column) { return &data[row * width + column]; }
__device__ __forceinline__ void set(int row, int column, half value) { data[row * width + column] = value; }
__device__ __forceinline__ void set_half2(int row, int column, half2 value) { ((half2*)data)[(row * width + column) / 2] = value; }
};
class MatrixView_q4_row
{
public:
const uint32_t* data;
const int height;
const int width;
__device__ __forceinline__ MatrixView_q4_row(const uint32_t* data, const int height, const int width)
: data(data), height(height), width(width)
{ }
__device__ __forceinline__ int item(int row, int column) const
{
int shift = (column & 0x07) * 4;
return (data[row * width / 8 + column / 8] >> shift) & 0x0f;
}
};
class MatrixView_q4_column
{
public:
const uint32_t* data;
const int height;
const int width;
__device__ __forceinline__ MatrixView_q4_column(const uint32_t* data, const int height, const int width)
: data(data), height(height), width(width)
{ }
__device__ __forceinline__ int item(int row, int column) const
{
int shift = (row & 0x07) * 4;
return (data[row / 8 * width + column] >> shift) & 0x0f;
}
__device__ __forceinline__ uint32_t item_uint32_t(int row, int column) { return data[row / 8 * width + column]; }
__device__ __forceinline__ const uint32_t* item_uint32_ptr(int row, int column) { return &data[row / 8 * width + column]; }
};
// TODO: Rewrite all these dot product functions using functors or something, move to q4_matmul.cu
// Accumulated dot product of 8-element row vectors in h and quantized column vectors in v, constant zero/scale
__device__ __forceinline__ half2 dot_product_8
(
const half2 acc,
const half2* h_ptr,
MatrixView_q4_column& v_,
const int v_row, // divisible by 8
const int v_column,
const half2 v_scale_2,
const uint32_t v_zero, // + 1 (!!)
const int count
)
{
const uint32_t* v_ptr = (const uint32_t*) v_.item_uint32_ptr(v_row, v_column);
half2 result = acc;
for (int i = 0; i < count; i++)
{
uint32_t v_read = *v_ptr; v_ptr += v_.width;
half v_0 = __int2half_rn((int)((v_read ) & 0x0f) - v_zero);
half v_1 = __int2half_rn((int)((v_read >> 4) & 0x0f) - v_zero);
half v_2 = __int2half_rn((int)((v_read >> 8) & 0x0f) - v_zero);
half v_3 = __int2half_rn((int)((v_read >> 12) & 0x0f) - v_zero);
half v_4 = __int2half_rn((int)((v_read >> 16) & 0x0f) - v_zero);
half v_5 = __int2half_rn((int)((v_read >> 20) & 0x0f) - v_zero);
half v_6 = __int2half_rn((int)((v_read >> 24) & 0x0f) - v_zero);
half v_7 = __int2half_rn((int)((v_read >> 28) ) - v_zero);
half2 v_01 = __halves2half2(v_0, v_1);
half2 v_23 = __halves2half2(v_2, v_3);
half2 v_45 = __halves2half2(v_4, v_5);
half2 v_67 = __halves2half2(v_6, v_7);
// half2 v_01 = q4_table[v_zero - 1][(v_read ) & 0xff]; // (constant memory is too slow apparently)
// half2 v_23 = q4_table[v_zero - 1][(v_read >> 8) & 0xff];
// half2 v_45 = q4_table[v_zero - 1][(v_read >> 16) & 0xff];
// half2 v_67 = q4_table[v_zero - 1][(v_read >> 24) ];
half2 tmp = __hmul2(*h_ptr++, v_01);
tmp = __hfma2(*h_ptr++, v_23, tmp);
tmp = __hfma2(*h_ptr++, v_45, tmp);
tmp = __hfma2(*h_ptr++, v_67, tmp);
result = __hfma2(v_scale_2, tmp, result);
}
return result;
}
__device__ __forceinline__ half dot_product_8_h
(
const half acc,
const half* h_ptr,
MatrixView_q4_column& v_,
const int v_row, // divisible by 8
const int v_column,
const half v_scale,
const uint32_t v_zero, // + 1 (!!)
const int count
)
{
const uint32_t* v_ptr = (const uint32_t*) v_.item_uint32_ptr(v_row, v_column);
half result = acc;
for (int i = 0; i < count; i++)
{
uint32_t v_read = *v_ptr; v_ptr += v_.width;
half v_0 = __int2half_rn((int)((v_read ) & 0x0f) - v_zero);
half v_1 = __int2half_rn((int)((v_read >> 4) & 0x0f) - v_zero);
half v_2 = __int2half_rn((int)((v_read >> 8) & 0x0f) - v_zero);
half v_3 = __int2half_rn((int)((v_read >> 12) & 0x0f) - v_zero);
half v_4 = __int2half_rn((int)((v_read >> 16) & 0x0f) - v_zero);
half v_5 = __int2half_rn((int)((v_read >> 20) & 0x0f) - v_zero);
half v_6 = __int2half_rn((int)((v_read >> 24) & 0x0f) - v_zero);
half v_7 = __int2half_rn((int)((v_read >> 28) ) - v_zero);
half tmp = __hmul(*h_ptr++, v_0);
tmp = __hfma(*h_ptr++, v_1, tmp);
tmp = __hfma(*h_ptr++, v_2, tmp);
tmp = __hfma(*h_ptr++, v_3, tmp);
tmp = __hfma(*h_ptr++, v_4, tmp);
tmp = __hfma(*h_ptr++, v_5, tmp);
tmp = __hfma(*h_ptr++, v_6, tmp);
tmp = __hfma(*h_ptr++, v_7, tmp);
result = __hfma(v_scale, tmp, result);
}
return result;
}
// Accumulated dot product of 8-element row vectors in h and quantized column vectors in v, constant zero/scale, with x_map
__device__ __forceinline__ half2 dot_product_8_x_map
(
const half2 acc,
MatrixView_half& h_,
const int h_row,
const int h_column, // divisible by 8
MatrixView_q4_column& v_,
const int v_row, // divisible by 8
const int v_column,
const half2 v_scale_2,
const uint32_t v_zero, // + 1 (!!)
const int count,
const uint32_t* x_map
)
{
const half* h_ptr = h_.item_ptr(h_row, 0);
const uint32_t* x_map_ptr = x_map + h_column;
const uint32_t* v_ptr = (const uint32_t*) v_.item_uint32_ptr(v_row, v_column);
half2 result = acc;
for (int i = 0; i < count; i++)
{
uint32_t v_read = *v_ptr; v_ptr += v_.width;
half v_0 = __int2half_rn((int)((v_read ) & 0x0f) - v_zero);
half v_1 = __int2half_rn((int)((v_read >> 4) & 0x0f) - v_zero);
half v_2 = __int2half_rn((int)((v_read >> 8) & 0x0f) - v_zero);
half v_3 = __int2half_rn((int)((v_read >> 12) & 0x0f) - v_zero);
half v_4 = __int2half_rn((int)((v_read >> 16) & 0x0f) - v_zero);
half v_5 = __int2half_rn((int)((v_read >> 20) & 0x0f) - v_zero);
half v_6 = __int2half_rn((int)((v_read >> 24) & 0x0f) - v_zero);
half v_7 = __int2half_rn((int)((v_read >> 28) ) - v_zero);
half2 v_01 = __halves2half2(v_0, v_1);
half2 v_23 = __halves2half2(v_2, v_3);
half2 v_45 = __halves2half2(v_4, v_5);
half2 v_67 = __halves2half2(v_6, v_7);
half h_0 = h_ptr[*x_map_ptr++];
half h_1 = h_ptr[*x_map_ptr++];
half h_2 = h_ptr[*x_map_ptr++];
half h_3 = h_ptr[*x_map_ptr++];
half h_4 = h_ptr[*x_map_ptr++];
half h_5 = h_ptr[*x_map_ptr++];
half h_6 = h_ptr[*x_map_ptr++];
half h_7 = h_ptr[*x_map_ptr++];
half2 h_01 = __halves2half2(h_0, h_1);
half2 h_23 = __halves2half2(h_2, h_3);
half2 h_45 = __halves2half2(h_4, h_5);
half2 h_67 = __halves2half2(h_6, h_7);
half2 tmp = __hmul2(h_01, v_01);
tmp = __hfma2(h_23, v_23, tmp);
tmp = __hfma2(h_45, v_45, tmp);
tmp = __hfma2(h_67, v_67, tmp);
result = __hfma2(v_scale_2, tmp, result);
}
return result;
}
__device__ __forceinline__ half dot_product_8_x_map_h
(
const half acc,
MatrixView_half& h_,
const int h_row,
const int h_column, // divisible by 8
MatrixView_q4_column& v_,
const int v_row, // divisible by 8
const int v_column,
const half v_scale,
const uint32_t v_zero, // + 1 (!!)
const int count,
const uint32_t* x_map
)
{
const half* h_ptr = h_.item_ptr(h_row, 0);
const uint32_t* x_map_ptr = x_map + h_column;
const uint32_t* v_ptr = (const uint32_t*) v_.item_uint32_ptr(v_row, v_column);
half result = acc;
for (int i = 0; i < count; i++)
{
uint32_t v_read = *v_ptr; v_ptr += v_.width;
half v_0 = __int2half_rn((int)((v_read ) & 0x0f) - v_zero);
half v_1 = __int2half_rn((int)((v_read >> 4) & 0x0f) - v_zero);
half v_2 = __int2half_rn((int)((v_read >> 8) & 0x0f) - v_zero);
half v_3 = __int2half_rn((int)((v_read >> 12) & 0x0f) - v_zero);
half v_4 = __int2half_rn((int)((v_read >> 16) & 0x0f) - v_zero);
half v_5 = __int2half_rn((int)((v_read >> 20) & 0x0f) - v_zero);
half v_6 = __int2half_rn((int)((v_read >> 24) & 0x0f) - v_zero);
half v_7 = __int2half_rn((int)((v_read >> 28) ) - v_zero);
half tmp = __hmul(h_ptr[*x_map_ptr++], v_0);
tmp = __hfma(h_ptr[*x_map_ptr++], v_1, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_2, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_3, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_4, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_5, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_6, tmp);
tmp = __hfma(h_ptr[*x_map_ptr++], v_7, tmp);
result = __hfma(v_scale, tmp, result);
}
return result;
}
#endif
|