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
|
#include "stdafx.h"
#include "tensorOpsTests.h"
#include "MlContext.h"
#include "TensorEx.h"
#include "../D3D/shaders.h"
#include "../D3D/Binder.h"
#include "testUtils.h"
#include "../Whisper/WhisperContext.h"
void DirectCompute::testMulMat( const ggml_tensor* src0, const ggml_tensor* src1, const ggml_tensor* dst, const void* tempBuffer )
{
return;
CaptureRaii capture;
const size_t nb00 = src0->nb[ 0 ];
const size_t nb01 = src0->nb[ 1 ];
if( src0->type != GGML_TYPE_F16 )
return; // TODO
if( nb01 < nb00 )
return; // TODO
WhisperContext& ctx = WhisperContext::current();
Tensor arg0, arg1;
check( arg0.create( *src0, eBufferUse::Immutable, true ) );
check( arg1.create( *src1, eBufferUse::Immutable, true ) );
TensorEx res;
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
ctx.mulMat( arg0, arg1, res );
std::vector<float> tv;
check( res.download( tv ) );
const size_t len = tv.size();
computeDiff( tv.data(), (const float*)dst->data, len ).print( "testMulMat-product" );
#if 0
dbgWriteBinaryFile( L"product-orig.bin", dst->data, len * 4 );
dbgWriteBinaryFile( L"product-gpu.bin", tv.data(), len * 4 );
__debugbreak();
#endif
}
#if 0
void DirectCompute::testMulMatReshape( const ggml_tensor* src1, const void* tempBuffer )
{
Tensor src;
check( src.create( *src1, eBufferUse::Immutable, true ) );
const size_t ne10 = (uint32_t)src1->ne[ 0 ];
const size_t ne11 = (uint32_t)src1->ne[ 1 ];
const size_t ne12 = (uint32_t)src1->ne[ 2 ];
const size_t ne13 = (uint32_t)src1->ne[ 3 ];
if( 1 != ne13 )
throw E_UNEXPECTED;
const size_t tempLength = ne10 * ne11 * ne12 * ne13;
Context& ctx = Context::current();
const ReadWriteViews& temp = ctx.temp.fp16( tempLength );
{
Binder bind;
ctx.cb.bind();
bindShader( eComputeShader::mulMatDotReshape );
ctx.cb.update( src );
bind.bind( src, temp );
context()->Dispatch( (UINT)ne11, (UINT)ne12, 1 );
}
std::vector<uint16_t> reshaped;
check( downloadBuffer( temp, reshaped ) );
computeDiff( reshaped.data(), (const uint16_t*)tempBuffer, reshaped.size() ).print( "testMulMatReshape" );
#if 0
dbgWriteBinaryFile( L"fp32.bin", src1->data, ggml_nbytes( src1 ) );
dbgWriteBinaryFile( L"fp16-cpu.bin", tempBuffer, reshaped.size() * 2 );
dbgWriteBinaryFile( L"fp16-gpu.bin", reshaped.data(), reshaped.size() * 2 );
__debugbreak();
#endif
}
#endif
void DirectCompute::computeMulMat( const ggml_tensor* src0, const ggml_tensor* src1, ggml_tensor* dst )
{
CaptureRaii capture;
const size_t nb00 = src0->nb[ 0 ];
const size_t nb01 = src0->nb[ 1 ];
if( src0->type != GGML_TYPE_F16 )
throw E_INVALIDARG;
if( nb01 < nb00 )
throw E_INVALIDARG;
WhisperContext& ctx = WhisperContext::current();
Tensor arg0, arg1;
check( arg0.create( *src0, eBufferUse::Immutable, true ) );
check( arg1.create( *src1, eBufferUse::Immutable, true ) );
TensorEx res;
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
ctx.mulMat( arg0, arg1, res );
check( res.download( dst->data ) );
}
void DirectCompute::testFlashAttention( const ggml_tensor* q, const ggml_tensor* k, const ggml_tensor* v, bool masked, const ggml_tensor* dst )
{
CaptureRaii capture;
Tensor Q, K, V;
TensorEx res;
check( Q.create( *q, eBufferUse::Immutable, true ) );
check( K.create( *k, eBufferUse::Immutable, true ) );
check( V.create( *v, eBufferUse::Immutable, true ) );
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
WhisperContext& ctx = WhisperContext::current();
ctx.flashAttention( Q, K, V, res, masked );
std::vector<float> tv;
check( res.download( tv ) );
const size_t len = tv.size();
computeDiff( tv.data(), (const float*)dst->data, len ).print( "testFlashAttention" );
}
void DirectCompute::computeFlashAttention( const ggml_tensor* q, const ggml_tensor* k, const ggml_tensor* v, bool masked, ggml_tensor* dst )
{
CaptureRaii capture;
Tensor Q, K, V;
TensorEx res;
check( Q.create( *q, eBufferUse::Immutable, true ) );
check( K.create( *k, eBufferUse::Immutable, true ) );
check( V.create( *v, eBufferUse::Immutable, true ) );
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
WhisperContext& ctx = WhisperContext::current();
ctx.flashAttention( Q, K, V, res, masked );
check( res.download( dst->data ) );
}
void DirectCompute::testConvolution( const ggml_tensor* src0, const ggml_tensor* src1, const ggml_tensor* dst )
{
CaptureRaii capture;
Tensor arg0, arg1;
check( arg0.create( *src0, eBufferUse::Immutable, true ) );
check( arg1.create( *src1, eBufferUse::Immutable, true ) );
TensorEx res;
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
WhisperContext& ctx = WhisperContext::current();
ctx.convolution( arg0, arg1, res );
std::vector<float> tv;
check( res.download( tv ) );
const size_t len = tv.size();
computeDiff( tv.data(), (const float*)dst->data, len ).print( "testConvolution" );
}
void DirectCompute::computeConvolution( const ggml_tensor* src0, const ggml_tensor* src1, ggml_tensor* dst )
{
CaptureRaii capture;
Tensor arg0, arg1;
check( arg0.create( *src0, eBufferUse::Immutable, true ) );
check( arg1.create( *src1, eBufferUse::Immutable, true ) );
TensorEx res;
check( res.create( *dst, eBufferUse::ReadWriteDownload, false ) );
WhisperContext& ctx = WhisperContext::current();
ctx.convolution( arg0, arg1, res );
res.download( dst->data );
}
|
