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
|
#include "stdafx.h"
#include "MlContext.h"
#include "../source/ggml.h"
#include "testUtils.h"
using namespace DirectCompute;
#define E_TYPE HRESULT_FROM_WIN32( ERROR_DATATYPE_MISMATCH )
static void dbgPrintSizeDiff( const char* what, __m128i ref, __m128i gpu )
{
std::array<int, 8> a;
_mm_storeu_si128( ( __m128i* ) & a[ 0 ], ref );
_mm_storeu_si128( ( __m128i* ) & a[ 4 ], gpu );
printf( "%s; reference [ %i, %i, %i, %i ], GPGPU [ %i, %i, %i, %i ]\n",
what,
a[ 0 ], a[ 1 ], a[ 2 ], a[ 3 ],
a[ 4 ], a[ 5 ], a[ 6 ], a[ 7 ] );
}
void MlContext::dbgPrintDifference( const ggml_tensor* reference, const Tensor& gpu, const char* what, bool trapToDebugger )
{
sTensorDiff diff;
const __m128i gpuSize = gpu.sizeVec();
const __m128i gpuStrides = gpu.stridesVec();
__m128i expectedStrides;
if( reference->type == GGML_TYPE_F32 )
{
if( gpu.getType() != eDataType::FP32 )
throw E_TYPE;
expectedStrides = _mm_slli_epi32( gpuStrides, 2 );
std::vector<float> v;
gpu.download( v );
diff = computeDiff( v.data(), (const float*)reference->data, v.size() );
}
else if( reference->type == GGML_TYPE_F16 )
{
if( gpu.getType() != eDataType::FP16 )
throw E_TYPE;
expectedStrides = _mm_slli_epi32( gpuStrides, 1 );
std::vector<uint16_t> v;
gpu.download( v );
diff = computeDiff( v.data(), (const uint16_t*)reference->data, v.size() );
}
else
throw E_NOTIMPL;
const __m128i ggmlSize = _mm_loadu_si128( ( const __m128i* ) & reference->ne[ 0 ] );
const __m128i ggmlStrides = _mm_loadu_si128( ( const __m128i* ) & reference->nb[ 0 ] );
if( !vectorEqual( gpuSize, ggmlSize ) )
dbgPrintSizeDiff( "dbgPrintDifference - size is different", ggmlSize, gpuSize );
// if( !vectorEqual( expectedStrides, ggmlStrides ) ) dbgPrintSizeDiff( "dbgPrintDifference - stride is different", ggmlStrides, expectedStrides );
diff.print( what );
if( trapToDebugger )
__debugbreak();
}
|
