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
290
291
292
|
#include "stdafx.h"
#include "MlContext.h"
#include "testUtils.h"
using namespace DirectCompute;
Tensor MlContext::createTensor( eDataType type, const std::array<uint32_t, 4>& ne )
{
Tensor res;
check( res.create( type, ne ) );
return res;
}
Tensor MlContext::createTensor( eDataType type, std::initializer_list<uint32_t> ne )
{
size_t nDims = ne.size();
if( 0 == nDims || nDims > 4 )
throw E_INVALIDARG;
std::array<uint32_t, 4> arr;
for( size_t i = 0; i < nDims; i++ )
arr[ i ] = ne.begin()[ i ];
for( size_t i = nDims; i < 4; i++ )
arr[ i ] = 1;
return createTensor( type, arr );
}
Tensor MlContext::conv_1d_1s( const Tensor& a, const Tensor& b )
{
assert( b.isMatrix() );
assert( a.ne[ 1 ] == b.ne[ 1 ] );
assert( a.ne[ 3 ] == 1 );
Tensor res = createTensor( eDataType::FP32, { b.ne[ 0 ], a.ne[ 2 ] } );
convolution( a, b, res );
return res;
}
Tensor MlContext::conv_1d_2s( const Tensor& a, const Tensor& b )
{
assert( b.isMatrix() );
assert( a.ne[ 1 ] == b.ne[ 1 ] );
assert( a.ne[ 3 ] == 1 );
Tensor res = createTensor( eDataType::FP32, { b.ne[ 0 ] / 2, a.ne[ 2 ] } );
#if 0
static PrintUniqueTensorSizes printSize( "conv_1d_2s" );
printSize.print( a, b );
#endif
convolution2( a, b, res );
return res;
}
namespace
{
inline bool canRepeat( const TensorShape& t0, const TensorShape& t1 )
{
return ( t1.ne[ 0 ] % t0.ne[ 0 ] == 0 ) &&
( t1.ne[ 1 ] % t0.ne[ 1 ] == 0 ) &&
( t1.ne[ 2 ] % t0.ne[ 2 ] == 0 ) &&
( t1.ne[ 3 ] % t0.ne[ 3 ] == 0 );
}
}
Tensor MlContext::cwiseBinary( const Tensor& a, const Tensor& b, eComputeShader cs )
{
assert( isSameShape( a, b ) );
Tensor res = createTensor( a.getType(), a.ne );
cwiseBinary( a, b, res, cs );
return res;
}
Tensor __declspec( noinline ) MlContext::view2d( const Tensor& a, uint32_t ne0, uint32_t ne1, uint32_t nb1, uint32_t offset )
{
if( 0 != offset )
throw E_NOTIMPL;
Tensor res = a;
res.ne = { ne0, ne1, 1, 1 };
res.nb[ 1 ] = nb1;
res.nb[ 2 ] = res.nb[ 3 ] = nb1 * ne1;
return res;
}
Tensor MlContext::transpose( const Tensor& a )
{
Tensor result;
// A magic number for _mm_shuffle_epi32 SSE2 instruction to swap two lower int32 lanes in a vector
constexpr int swapXy = _MM_SHUFFLE( 3, 2, 0, 1 );
__m128i v = a.sizeVec();
v = _mm_shuffle_epi32( v, swapXy );
store( result.ne, v );
v = a.stridesVec();
v = _mm_shuffle_epi32( v, swapXy );
store( result.nb, v );
result.setGpuViews( a, a );
return result;
}
Tensor MlContext::norm( const Tensor& a )
{
Tensor res = createTensor( a.getType(), a.ne );
norm( a, res );
return res;
}
Tensor MlContext::mulMat( const Tensor& a, const Tensor& b )
{
if( !canMulMat( a, b ) )
throw E_INVALIDARG;
Tensor res = createTensor( eDataType::FP32, { a.ne[ 1 ], b.ne[ 1 ], a.ne[ 2 ], b.ne[ 3 ] } );
if constexpr( enableInexactOptimizations )
mulMatTiled( a, b, res );
else
mulMat( a, b, res );
#if 0
Tensor testTiled;
check( testTiled.create( eDataType::FP32, res.ne ) );
mulMatTiled( a, b, testTiled );
std::vector<float> current, tiled;
res.download( current );
testTiled.download( tiled );
sTensorDiff diff = computeDiff( current.data(), tiled.data(), current.size() );
diff.print( "mulMatTiled" );
#endif
return res;
}
Tensor MlContext::mulMatEx( const Tensor& a, const Tensor& b, const char* tagName )
{
if( !canMulMat( a, b ) )
throw E_INVALIDARG;
if( 0 != a.nb[ 0 ] )
throw E_INVALIDARG; // The first argument is expected to be pre-transposed
const uint16_t tag = profiler.setNextTag( tagName );
if( b.ne[ 1 ] != 1 )
{
if( b.nb[ 0 ] != 0 )
{
Tensor rhs = reshapePanels( b );
profiler.setNextTag( tag );
return mulMatTiledEx( a, rhs );
}
else
{
// Second argument already reshaped into these panels
return mulMatTiledEx( a, b );
}
}
else
{
if( 0 != b.nb[ 0 ] )
return mulMatByRowTiledEx( a, b );
// That shader requires classic VRAM layout of the second argument, gonna fail with pre-transposed one
throw E_INVALIDARG;
}
}
Tensor MlContext::permute( const Tensor& a, uint8_t axis0, uint8_t axis1, uint8_t axis2, uint8_t axis3 )
{
assert( axis0 < 4 );
assert( axis1 < 4 );
assert( axis2 < 4 );
assert( axis3 < 4 );
assert( axis0 != axis1 );
assert( axis0 != axis2 );
assert( axis0 != axis3 );
assert( axis1 != axis2 );
assert( axis1 != axis3 );
assert( axis2 != axis3 );
Tensor res = a;
res.ne[ axis0 ] = a.ne[ 0 ];
res.ne[ axis1 ] = a.ne[ 1 ];
res.ne[ axis2 ] = a.ne[ 2 ];
res.ne[ axis3 ] = a.ne[ 3 ];
res.nb[ axis0 ] = a.nb[ 0 ];
res.nb[ axis1 ] = a.nb[ 1 ];
res.nb[ axis2 ] = a.nb[ 2 ];
res.nb[ axis3 ] = a.nb[ 3 ];
return res;
}
Tensor MlContext::flashAttention( const Tensor& q, const Tensor& k, const Tensor& v, bool masked )
{
if( !canMulMat( k, q ) )
throw E_INVALIDARG;
if constexpr( enableInexactOptimizations )
{
if( !masked )
{
profiler.setNextTag( "flashAttn.1" );
Tensor tmp = mulMat( k, q );
profiler.setNextTag( "flashAttention" );
const float tempScale = (float)( 1.0 / sqrt( (double)(int)q.ne[ 0 ] ) );
softMax( tmp, tempScale );
profiler.setNextTag( "flashAttn.2" );
return mulMat( v, tmp );
}
}
Tensor res = createTensor( eDataType::FP32, q.ne );
flashAttention( q, k, v, res, masked );
#if 0
Tensor tmpMat = mulMat( k, q );
float scale = (float)( 1.0 / sqrt( (double)(int)q.ne[ 0 ] ) );
softMax( tmpMat, scale );
Tensor testRes = mulMat( v, tmpMat );
computeDiff( res, testRes ).print( "flashAttention mulmat" );
#endif
return res;
}
Tensor MlContext::copy( const Tensor& a, eDataType type, std::initializer_list<uint32_t> size )
{
const size_t dims = size.size();
if( 0 == dims || dims > 4 )
throw E_BOUNDS;
size_t nRequested = 1;
for( size_t i = 0; i < dims; i++ )
{
uint32_t n = size.begin()[ i ];
nRequested *= n;
}
if( nRequested != a.countElements() )
throw E_INVALIDARG;
const eDataType st = a.getType();
Tensor res;
if( a.isContinuous() && st == type )
{
// Same type, and it's dense - no need to call any compute shaders, equal to reshape
res = a;
for( size_t i = 0; i < dims; i++ )
res.ne[ i ] = size.begin()[ i ];;
for( size_t i = dims; i < 4; i++ )
res.ne[ i ] = 1;
res.setDenseStrides();
}
else
{
// Either converting non-continuous to continuous, or converting types
res = createTensor( type, size );
copyImpl( a, res, st == eDataType::FP32 && type == eDataType::FP16 );
}
return res;
}
void MlContext::copyInPlace( Tensor& dest, const Tensor& a, eDataType type, std::initializer_list<uint32_t> size )
{
assert( type == dest.getType() );
const size_t dims = size.size();
if( 0 == dims || dims > 4 )
throw E_BOUNDS;
size_t nRequested = 1;
for( size_t i = 0; i < dims; i++ )
{
uint32_t n = size.begin()[ i ];
nRequested *= n;
}
if( nRequested != a.countElements() || nRequested != dest.countElements() )
throw E_INVALIDARG;
// Reshape the destination
for( size_t i = 0; i < dims; i++ )
dest.ne[ i ] = size.begin()[ i ];
for( size_t i = dims; i < 4; i++ )
dest.ne[ i ] = 1;
dest.setDenseStrides();
// Call the shader
const eDataType st = a.getType();
copyImpl( a, dest, st == eDataType::FP32 && type == eDataType::FP16 );
}
|
