static void Init_b(long long * gradIn, long long * gradOut, long long * primitives)
{
dnnError_t err;
//gradOut, layout is user or mkl
dnnLayout_t lt_out_b = (dnnLayout_t)primitives[POOL_L_B_O];
dnnLayout_t lt_out = (dnnLayout_t)gradOut[MKLLayout];
if (lt_out==NULL) lt_out = (dnnLayout_t)primitives[POOL_L_O];
//create conversion and buff if necessary
dnnPrimitive_t cv_out_b = NULL; float * buf_out_b = NULL;
CHECK_ERR( try_convert(&cv_out_b, &buf_out_b, lt_out, lt_out_b) , err );
//save
primitives[CV_POOLING_BACKWARD_OUTPUT] = (long long)cv_out_b;
primitives[BUFFER_POOLING_BACKWARD_OUTPUT] = (long long)buf_out_b;
//gradIn, layout
gradIn[CPULayout] = primitives[POOL_L_I];
dnnLayout_t lt_in_b = (dnnLayout_t)primitives[POOL_L_B_I];
float* buf_in_b = NULL;
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buf_in_b), lt_in_b), err );
primitives[BUFFER_POOLING_BACKWARD_INPUT] = (long long)buf_in_b;
ERR_RETURN:
return;
}
//gradOut: output gradient of CONV layer, known parameters
//gradIn: input gradient, to be calculated
static void Conv_bdata_init(
long long * gradIn,
long long * gradOut,
int N, int oC, int oH, int oW,
long long * weight,
long long * primitives)
{
dnnError_t err;
//get gradOut layout, create conversion if necessary
dnnLayout_t lt_out = (dnnLayout_t)gradOut[MKLLayout];
if(lt_out==NULL)
{
lt_out = (dnnLayout_t)primitives[L_O];
}
dnnPrimitive_t cv_out_bdata = NULL; float * buf_out_bdata = NULL;
CHECK_ERR( try_convert(&cv_out_bdata, &buf_out_bdata, lt_out, (dnnLayout_t)primitives[L_BD_O]) , err );
primitives[CONVERT_BWDDATA_OUTPUT] = (long long)cv_out_bdata;
primitives[BUFFER_BWDDATA_OUTPUT] = (long long)buf_out_bdata;
//for filter
dnnLayout_t lt_filter = (dnnLayout_t)primitives[L_W];
dnnLayout_t lt_filter_bdata = (dnnLayout_t)primitives[L_BD_W];
dnnPrimitive_t cv_filter_bdata = NULL; float * buf_filter_bdata = NULL;
CHECK_ERR( try_convert(&cv_filter_bdata, &buf_filter_bdata, lt_filter, lt_filter_bdata), err );
primitives[BUFFER_BWDDATA_FILTER] = (long long)buf_filter_bdata;
primitives[CONVERT_BWDDATA_FILTER] = (long long)cv_filter_bdata;
//create gradInput layout and memory
dnnLayout_t lt_in_bdata = (dnnLayout_t)primitives[L_BD_I];
float * buf_in_bdata = (float*)(gradIn[CPUPtr]);
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buf_in_bdata), lt_in_bdata), err );
primitives[BUFFER_BWDDATA_INPUT] = (long long)buf_in_bdata;
gradIn[CPULayout] = (long long)(dnnLayout_t)primitives[L_I_CHWN];
float* gradOutTransPtr = NULL;
if (gradOut[MKLLayout] == 0)
{
gradOutTransPtr = (float*)malloc(N*oC*oH*oW*sizeof(float));
}
primitives[BUFFER_TRANS_OUTPUT] = (long long)gradOutTransPtr;
ERR_RETURN:
return;
}
void MaxPooling_bprop(
unsigned long long gradOutput, //input, N*outC*outH*outW
unsigned long long gradInput, //output result
unsigned long long dnnprimitives,
int initOK, const float beta)
{
dnnError_t err;
long long* primitives = (long long*)dnnprimitives;
if (initOK == 0)
{
Init_b((long long *)gradInput, (long long *)gradOutput, primitives);
}
//get resource
float* resPool[dnnResourceNumber] = {0};
float* OutPtr= GetPtr(gradOutput);
resPool[dnnResourceDiffSrc] = (float*)primitives[BUFFER_POOLING_BACKWARD_INPUT];
resPool[dnnResourceDiffDst] = OutPtr;
resPool[dnnResourceWorkspace] = (float*)primitives[BUFFER_POOLING_FORWARD_WORKSPACE];
//make conversion for gradeOut if necessary
dnnPrimitive_t cv_out_b = (dnnPrimitive_t)(primitives[CV_POOLING_BACKWARD_OUTPUT]);
if (cv_out_b)
{
float* buf_out_b = (float*)primitives[BUFFER_POOLING_BACKWARD_OUTPUT];
CHECK_ERR( dnnConversionExecute_F32(cv_out_b, OutPtr, buf_out_b), err );
resPool[dnnResourceDiffDst] = buf_out_b;
}
long long grad_in_len = (long long)dnnLayoutGetMemorySize_F32((dnnLayout_t)primitives[POOL_L_B_I]) ;
float * tempPtr = (float*)primitives[BUFFER_POOLING_BACKWARD_INPUT];
#pragma omp parallel for
for (long long i = 0; i < grad_in_len/4; ++i)
{
tempPtr[i] = 0;
}
CHECK_ERR( dnnExecute_F32((dnnPrimitive_t)primitives[POOLING_BACKWARD], (void**)resPool), err );
if(beta != 0.0)
{
//require to add previous delta
long long* ptr_gradInput = (long long*)gradInput;
float* pFirstBuf = GetPtr(gradInput);
dnnLayout_t layout_pre_delta = (dnnLayout_t)ptr_gradInput[MKLLayout];
if(layout_pre_delta == NULL) layout_pre_delta = (dnnLayout_t)primitives[POOL_L_I];
dnnLayout_t layout_add_delta = (dnnLayout_t)primitives[POOL_L_B_I];
float* temp_memory = NULL;
if (!dnnLayoutCompare_F32(layout_add_delta, layout_pre_delta))
{
CHECK_ERR( dnnAllocateBuffer_F32((void**)&temp_memory, layout_add_delta) , err );
dnnPrimitive_t cv = NULL;
CHECK_ERR( dnnConversionCreate_F32(&cv, layout_pre_delta, layout_add_delta), err );
CHECK_ERR( dnnConversionExecute_F32(cv, pFirstBuf, temp_memory), err );
pFirstBuf = temp_memory;
}
long len = (long long)dnnLayoutGetMemorySize_F32(layout_add_delta) / 4 ;
cblas_saxpy(len, 1.0, pFirstBuf, 1, (float*)primitives[BUFFER_POOLING_BACKWARD_INPUT], 1);
if (temp_memory != NULL)
dnnReleaseBuffer_F32(temp_memory);
}
((long long *)gradInput)[MKLLayout] = primitives[POOL_L_B_I];
((long long *)gradInput)[MKLPtr] = primitives[BUFFER_POOLING_BACKWARD_INPUT];
ERR_RETURN:
return;
}
//useMaxPooling, or averagePooling
//useCaffe use ceil mode for pooling output dim
static void Init_f(
long long * input,
long long * output,
long long * primitives,
int N, int inC, int inH, int inW,
int kH, int kW, int dH, int dW,
int padH,int padW,
int outC, int outH,int outW,
int useMaxPooling,
int useCaffe)
{
dnnError_t err;
//dimension
size_t inputSize[DIM4] = { inW, inH, inC, N};
size_t outputSize[DIM4] = {outW, outH, outC, N};
size_t inputStrides1[DIM4] = {1, inW, inW*inH, inW*inH*inC};
size_t outputStrides1[DIM4] = {1, outW, outW*outH, outW*outH*outC};
//CHWN
size_t inputStrides[DIM4] = {N, N*inW, N*inW*inH, 1};
size_t outputStrides[DIM4] = {N, N*outW, N*outW*outH, 1};
size_t kernelSize[2] = { kW, kH};
size_t kernelStride[2] = { dW, dH};
//calculate pad
int padH2 = (outH-1)*dH + kH - inH - padH;
int padW2 = (outW-1)*dW + kW - inW - padW;
int symm = 0;
if (padH2==padH && padW2==padW) symm = 1;
if (padH2<0) padH2 = 0;
if (padW2<0) padW2 = 0;
int pad_dim4[DIM4] = {-padW, -padH, -padW2,-padH2};
int pad_dim2[DIM2] = {-padW, -padH};
int inputOffset[DIM2] = { 0, 0};
//create user layout
dnnLayout_t lt_out = NULL, lt_in = NULL;
CHECK_ERR( dnnLayoutCreate_F32(<_in, DIM4, inputSize, inputStrides) , err );
CHECK_ERR( dnnLayoutCreate_F32(<_out, DIM4, outputSize, outputStrides) , err );
primitives[POOL_L_I] = (long long)lt_in;
primitives[POOL_L_O] = (long long)lt_out;
//create MKL input layout
dnnLayout_t lt_in_f = (dnnLayout_t)input[MKLLayout];
if(lt_in_f==NULL)
{
lt_in_f = lt_in;
}
primitives[POOL_L_F_I] = (long long)lt_in_f;
//create operation
dnnPrimitive_t pool_f = NULL, pool_b = NULL;
dnnPrimitiveAttributes_t attributes = NULL;
CHECK_ERR( dnnPrimitiveAttributesCreate_F32(&attributes), err );
if (useMaxPooling==1)
{
if(useCaffe || symm)
{
CHECK_ERR( dnnPoolingCreateForward_F32 (&pool_f, attributes, dnnAlgorithmPoolingMax,lt_in_f, kernelSize, kernelStride, pad_dim2, dnnBorderZeros), err );
CHECK_ERR( dnnPoolingCreateBackward_F32(&pool_b, attributes, dnnAlgorithmPoolingMax,lt_in_f, kernelSize, kernelStride, pad_dim2, dnnBorderZeros), err );
}
else
{
CHECK_ERR( dnnPoolingCreateForward_F32 (&pool_f, attributes, dnnAlgorithmPoolingMax,lt_in_f, kernelSize, kernelStride, pad_dim4, dnnBorderZerosAsymm), err );
CHECK_ERR( dnnPoolingCreateBackward_F32(&pool_b, attributes, dnnAlgorithmPoolingMax,lt_in_f, kernelSize, kernelStride, pad_dim4, dnnBorderZerosAsymm), err );
}
}
else
{
if(useCaffe || symm)
{
CHECK_ERR( dnnPoolingCreateForward_F32 (&pool_f, attributes, dnnAlgorithmPoolingAvg,lt_in_f, kernelSize, kernelStride, pad_dim2, dnnBorderZeros), err );
CHECK_ERR( dnnPoolingCreateBackward_F32(&pool_b, attributes, dnnAlgorithmPoolingAvg,lt_in_f, kernelSize, kernelStride, pad_dim2, dnnBorderZeros), err );
}
else
{
CHECK_ERR( dnnPoolingCreateForward_F32 (&pool_f, attributes, dnnAlgorithmPoolingAvg,lt_in_f, kernelSize, kernelStride, pad_dim4, dnnBorderZerosAsymm), err );
CHECK_ERR( dnnPoolingCreateBackward_F32(&pool_b, attributes, dnnAlgorithmPoolingAvg,lt_in_f, kernelSize, kernelStride, pad_dim4, dnnBorderZerosAsymm), err );
}
}
primitives[POOLING_FORWARD] = (long long)pool_f;
primitives[POOLING_BACKWARD] = (long long)pool_b;
//create mkl layout for output
dnnLayout_t lt_out_f = NULL, lt_out_b = NULL, lt_in_b = NULL;
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_f, pool_f, dnnResourceDst), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_in_b, pool_f, dnnResourceSrc), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_b, pool_f, dnnResourceDst), err );
primitives[POOL_L_F_O] = (long long)lt_out_f;
primitives[POOL_L_B_I] = (long long)lt_in_b;
primitives[POOL_L_B_O] = (long long)lt_out_b;
//create work space , to record max location?
dnnLayout_t lt_space = NULL; float* buf_space = NULL;
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_space, pool_f, dnnResourceWorkspace), err );
CHECK_ERR( dnnAllocateBuffer_F32((void**)&buf_space, lt_space) , err );
primitives[BUFFER_POOLING_FORWARD_WORKSPACE] = (long long)buf_space;
//output layout
output[CPULayout] = (long long)lt_out;
float* buf_out_f = NULL;
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buf_out_f), lt_out_f), err );
primitives[BUFFER_POOLING_FORWARD_OUTPUT] = (long long)buf_out_f;
ERR_RETURN:
return;
}
static void THNN_(BatchNormalization_MKLDNN_init_forward)(
THLongTensor *primitives,
int N,
int inC,
int inH,
int inW,
double eps)
{
dnnError_t err;
dnnPrimitive_t bn_forward = NULL;
dnnPrimitive_t bn_backward = NULL;
dnnPrimitive_t bn_bwd_scaleshift = NULL;
size_t inputSize[dimension] = {inW,inH,inC,N};
size_t inputStrides[dimension] = { 1, inW, inH * inW, inC * inH * inW };
dnnLayout_t lt_user_input = NULL;
if(primitives->storage->data[BN_LAYOUT_INPUT] == 0)
{
CHECK_ERR( dnnLayoutCreate_F32(<_user_input, dimension, inputSize, inputStrides) , err );
#if CONVERSION_LOG
fprintf(stderr ,"MKLDNN BN get input layout FAIL......\n");
#endif
}
else{
lt_user_input = (dnnLayout_t)primitives->storage->data[BN_LAYOUT_INPUT];
#if CONVERSION_LOG
fprintf(stderr ,"MKLDNN BN get input layout OK\n");
#endif
}
CHECK_ERR( dnnBatchNormalizationCreateForward_F32(&bn_forward,NULL,lt_user_input,eps), err );
CHECK_ERR( dnnBatchNormalizationCreateBackwardData_F32(&bn_backward,NULL,lt_user_input,eps), err );
CHECK_ERR( dnnBatchNormalizationCreateBackwardScaleShift_F32(&bn_bwd_scaleshift,NULL,lt_user_input,eps), err );
dnnLayout_t lt_bn_forward_workspace,lt_bn_forward_scaleshift,lt_bn_forward_output,lt_bn_backward_input;
real * buffer_forward_workspace = NULL; real * buffer_forward_scaleshift = NULL;
real * buffer_forward_output = NULL; real * buffer_backward_input = NULL;
dnnLayoutCreateFromPrimitive_F32(<_bn_forward_workspace, bn_forward, dnnResourceWorkspace);
dnnLayoutCreateFromPrimitive_F32(<_bn_forward_output, bn_forward, dnnResourceDst);
dnnLayoutCreateFromPrimitive_F32(<_bn_forward_scaleshift, bn_forward, dnnResourceScaleShift);
dnnLayoutCreateFromPrimitive_F32(<_bn_backward_input, bn_backward, dnnResourceDiffSrc);
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_forward_workspace), lt_bn_forward_workspace), err );
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_forward_scaleshift), lt_bn_forward_scaleshift), err );
//CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_forward_output), lt_bn_forward_output), err );
//CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_backward_input), lt_bn_backward_input), err );
int size1 = dnnLayoutGetMemorySize_F32(lt_bn_forward_output);
int size2 = inW*inH*inC*N*4;
if(size1 == size2)
{
#if CONVERSION_LOG
fprintf(stderr ,"MKLDNN BN forward ouput layout match OK\n");
#endif
}
else
{
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_forward_output), lt_bn_forward_output), err );
fprintf(stderr ,"MKLDNN BN forward ouput layout match FAIL, size1 = %d, size2 = %d \n", size1, size2);
}
size1 = dnnLayoutGetMemorySize_F32(lt_bn_backward_input);
if(size1 == size2)
{
#if CONVERSION_LOG
fprintf(stderr ,"MKLDNN MaxPooling bwddata input layout match OK\n");
#endif
}
else
{
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buffer_backward_input), lt_bn_backward_input), err );
fprintf(stderr ,"MKLDNN MaxPooling bwddata input layout match FAIL, size1 = %d, size2 = %d \n", size1, size2);
}
//save the dnnPrimitive to THTensor(long int array)
primitives->storage->data[BN_LAYOUT_FORWARD_OUTPUT] = (long long)lt_bn_forward_output;
primitives->storage->data[BN_LAYOUT_BACKWARD_INPUT] = (long long)lt_bn_backward_input;
primitives->storage->data[BN_FORWARD] = (long long)bn_forward;
primitives->storage->data[BN_BACKWARD] = (long long)bn_backward;
primitives->storage->data[BN_SCALESHIFT] = (long long)bn_bwd_scaleshift;
primitives->storage->data[BUFFER_BN_FORWARD_WORKSPACE] = (long long)buffer_forward_workspace;
primitives->storage->data[BUFFER_BN_FORWARD_SCALESHIFT] = (long long)buffer_forward_scaleshift;
primitives->storage->data[BUFFER_BN_FORWARD_OUTPUT] = (long long)buffer_forward_output;
primitives->storage->data[BUFFER_BN_BACKWARD_INPUT] = (long long)buffer_backward_input;
primitives->storage->data[BUFFER_BN_BACKWARD_WORKSPACE] = (long long)buffer_forward_workspace;
}
static void Conv_bfilter_init(
long long * input,
long long * gradOutput,
long long * gradWeight,
long long * primitives,
int N, int oC, int oH, int oW)
{
dnnError_t err;
//for gradOut
dnnLayout_t lt_out = (dnnLayout_t)(gradOutput[MKLLayout]);
if(lt_out==NULL) lt_out = (dnnLayout_t)primitives[L_O];
dnnPrimitive_t cv_out_bfilter = NULL; float* buf_out_bfilter = NULL;
CHECK_ERR( try_convert(&cv_out_bfilter, &buf_out_bfilter, lt_out, (dnnLayout_t)primitives[L_BF_O]) , err );
primitives[CONVERT_BWDFILTER_OUTPUT] = (long long)cv_out_bfilter;
primitives[BUFFER_BWDFILTER_OUTPUT] = (long long)buf_out_bfilter;
//for the first layer without delta, input gradOut should first be transposed
float* gradOutTransPtr = NULL;
if ( gradOutput[MKLLayout] == 0 && primitives[BUFFER_TRANS_OUTPUT] == 0)
{
gradOutTransPtr = (float*)malloc(N*oC*oH*oW*sizeof(float));
primitives[BUFFER_TRANS_OUTPUT] = (long long)gradOutTransPtr;
}
//for filter
dnnLayout_t lt_filter = (dnnLayout_t)primitives[L_W];
dnnLayout_t lt_filter_bfilter = (dnnLayout_t)primitives[L_BF_W];
dnnPrimitive_t cv_filter_bfilter = NULL; float * buf_filter_bfilter = NULL;
if(!dnnLayoutCompare_F32(lt_filter_bfilter, lt_filter))
{
CHECK_ERR( dnnConversionCreate_F32(&cv_filter_bfilter, lt_filter_bfilter, lt_filter), err);
CHECK_ERR( dnnAllocateBuffer_F32((void**)&buf_filter_bfilter, lt_filter_bfilter), err);
}
primitives[BUFFER_BWDFILTER_FILTER] = (long long)buf_filter_bfilter;
primitives[CONVERT_BWDFILTER_FILTER] = (long long)cv_filter_bfilter;
//for input
dnnLayout_t lt_in_real = (dnnLayout_t)input[MKLLayout];
if(lt_in_real==NULL)
{
lt_in_real = (dnnLayout_t)primitives[L_I];
}
dnnLayout_t lt_in_bfilter = (dnnLayout_t)primitives[L_BF_I];
dnnPrimitive_t cv_in_bfilter = NULL; float* buf_in_bfilter = (float*)(input[CPUPtr]);
CHECK_ERR( try_convert(&cv_in_bfilter, &buf_in_bfilter, lt_in_real, lt_in_bfilter), err );
primitives[BUFFER_BWDFILTER_INPUT] = (long long)buf_in_bfilter;
primitives[CONVERT_BWDFILTER_INPUT] = (long long)cv_in_bfilter;
//if has bias
if (primitives[BDW_BIAS_INDEX] != 0)
{
//convert for grad_bias if necessary
dnnLayout_t lt_bias_bias = (dnnLayout_t)primitives[L_B_B];
dnnLayout_t lt_bias = (dnnLayout_t)primitives[L_B];
dnnPrimitive_t cv_bias_bias = NULL; float * buf_bias_bias = NULL;
CHECK_ERR( dnnConversionCreate_F32(&cv_bias_bias, lt_bias_bias, lt_bias), err);
CHECK_ERR( dnnAllocateBuffer_F32((void**)&buf_bias_bias, lt_bias_bias), err);
primitives[BUFFER_BIAS_BIAS] = (long long)buf_bias_bias;
primitives[CV_BIAS_BIAS] = (long long)cv_bias_bias;
//convert for grad_out if necessary
dnnLayout_t lt_bias_out = (dnnLayout_t)primitives[L_B_O];
dnnPrimitive_t cv_out_bias = NULL; float* buf_out_bias = (float*)(input[CPUPtr]);
CHECK_ERR( try_convert(&cv_out_bias, &buf_out_bias, lt_out, lt_bias_out), err );
primitives[BUFFER_BIAS_OUT] = (long long)buf_out_bias;
primitives[CV_BIAS_OUT] = (long long)cv_out_bias;
}
ERR_RETURN:
return;
}
static int Conv_f_init(
long long * input,
long long * output,
long long * weight,
long long * primitives,
int N, int inC, int inH, int inW,
int kH, int kW,
int dH, int dW,
int padH, int padW,
int outC, int outH, int outW,
int hasBias)
{
dnnError_t err;
//init dimensions
size_t inputSize[DIM4] = { inW, inH, inC, N};
size_t outputSize[DIM4] = {outW, outH, outC, N};
size_t filterSize[DIM4] = { kW, kH, inC, outC};
size_t stride[DIM2] = { dW, dH};
int pad[DIM2] = {-padW, -padH};
size_t biasSize[1] = {outC};
size_t biasStrides[1] = { 1 };
//using NCHW layout
size_t filterStridesNCHW[DIM4] = {1, kW, kW*kH, kW*kH*inC};
size_t inputStridesNCHW[DIM4] = {1, inW, inW*inH, inW*inH*inC};
size_t outputStridesNCHW[DIM4] = {1, outW, outW*outH, outW*outH*outC};
//CHWN
size_t filterStridesCHWN[DIM4] = {outC, outC*kW, outC*kW*kH, 1};
size_t inputStridesCHWN[DIM4] = {N, N*inW, N*inW*inH, 1};
size_t outputStridesCHWN[DIM4] = {N, N*outW, N*outW*outH, 1};
//create execute and save into primitives
dnnPrimitiveAttributes_t attributes = NULL;
CHECK_ERR( dnnPrimitiveAttributesCreate_F32(&attributes), err );
dnnPrimitive_t conv_f = NULL; //forward operation
dnnPrimitive_t conv_bdata = NULL; //backward calculate gradient input
dnnPrimitive_t conv_bfilter = NULL; //backward calculate gradient filter(weight)
dnnPrimitive_t conv_b_bias = NULL; //backward bias
//create layout and save
//lt_in, layout of input in NCHW form
//lt_filter_f, required layout (MKL layout) for forward for weight
//lt_out_bfilter, required layout for backward weight update for output
dnnLayout_t lt_in_NCHW, lt_filter, lt_out_NCHW, lt_in_CHWN, lt_out_CHWN, lt_bias_CHWN=NULL;
dnnLayout_t lt_in_f, lt_filter_f, lt_out_f, lt_bias_f;
dnnLayout_t lt_in_bdata, lt_filter_bdata, lt_out_bdata, lt_bias_bdata;
dnnLayout_t lt_in_bfilter, lt_filter_bfilter, lt_out_bfilter,lt_bias_bias, lt_out_bias;
if (hasBias)
{
CHECK_ERR(dnnConvolutionCreateForwardBias_F32( &conv_f, attributes, dnnAlgorithmConvolutionDirect, DIM4, inputSize, outputSize, filterSize, stride, pad, dnnBorderZeros),err);
CHECK_ERR(dnnConvolutionCreateForwardBias_F32( &conv_f, attributes, dnnAlgorithmConvolutionDirect, DIM4, inputSize, outputSize, filterSize, stride, pad, dnnBorderZeros),err);
CHECK_ERR( dnnLayoutCreate_F32(<_bias_CHWN, 1, biasSize, biasStrides), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_bias_f, conv_f, dnnResourceBias ) , err );
CHECK_ERR(dnnConvolutionCreateBackwardBias_F32( &conv_b_bias, attributes, dnnAlgorithmConvolutionDirect, DIM4, outputSize),err);
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_bias_bias, conv_b_bias, dnnResourceDiffBias) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_bias, conv_b_bias, dnnResourceDiffDst) , err );
}
else
CHECK_ERR(dnnConvolutionCreateForward_F32( &conv_f, attributes, dnnAlgorithmConvolutionDirect, DIM4, inputSize, outputSize, filterSize, stride, pad, dnnBorderZeros),err);
CHECK_ERR(dnnConvolutionCreateBackwardData_F32( &conv_bdata, attributes, dnnAlgorithmConvolutionDirect, DIM4, inputSize, outputSize, filterSize, stride, pad, dnnBorderZeros),err);
CHECK_ERR(dnnConvolutionCreateBackwardFilter_F32(&conv_bfilter, attributes, dnnAlgorithmConvolutionDirect, DIM4, inputSize, outputSize, filterSize, stride, pad, dnnBorderZeros),err);
primitives[FORWARD_INDEX] = (long long)conv_f;
primitives[BWD_DATA_INDEX] = (long long)conv_bdata;
primitives[BWD_FILTER_INDEX] = (long long)conv_bfilter;
primitives[BDW_BIAS_INDEX] = (long long)conv_b_bias;
CHECK_ERR( dnnLayoutCreate_F32(<_in_NCHW, DIM4, inputSize, inputStridesNCHW), err );
CHECK_ERR( dnnLayoutCreate_F32(<_in_CHWN, DIM4, inputSize, inputStridesCHWN), err );
CHECK_ERR( dnnLayoutCreate_F32(<_filter, DIM4, filterSize, filterStridesCHWN), err );
CHECK_ERR( dnnLayoutCreate_F32(<_out_NCHW, DIM4, outputSize, outputStridesNCHW), err );
CHECK_ERR( dnnLayoutCreate_F32(<_out_CHWN, DIM4, outputSize, outputStridesCHWN), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_in_f, conv_f, dnnResourceSrc ) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_filter_f, conv_f, dnnResourceFilter), err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_f, conv_f, dnnResourceDst ) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_in_bdata, conv_bdata, dnnResourceDiffSrc) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_filter_bdata, conv_bdata, dnnResourceFilter) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_bdata, conv_bdata, dnnResourceDiffDst) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_in_bfilter, conv_bfilter, dnnResourceSrc) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_filter_bfilter, conv_bfilter, dnnResourceDiffFilter) , err );
CHECK_ERR( dnnLayoutCreateFromPrimitive_F32(<_out_bfilter, conv_bfilter, dnnResourceDiffDst) , err );
//here assume NCHW (CHWN will be transposed)
primitives[L_I] = (long long)lt_in_NCHW;
primitives[L_O] = (long long)lt_out_NCHW;
primitives[L_W] = (long long)lt_filter;
primitives[L_B] = (long long)lt_bias_CHWN;
primitives[L_F_I] = (long long)lt_in_f;
primitives[L_F_O] = (long long)lt_out_f;
primitives[L_F_W] = (long long)lt_filter_f;
primitives[L_F_B] = (long long)lt_bias_f;
primitives[L_BD_I] = (long long)lt_in_bdata;
primitives[L_BD_O] = (long long)lt_out_bdata;
primitives[L_BD_W] = (long long)lt_filter_bdata;
primitives[L_BF_I] = (long long)lt_in_bfilter;
primitives[L_BF_O] = (long long)lt_out_bfilter;
primitives[L_BF_W] = (long long)lt_filter_bfilter;
primitives[L_I_CHWN] = (long long)lt_in_CHWN;
primitives[L_O_CHWN] = (long long)lt_out_CHWN;
primitives[L_B_B] = (long long)lt_bias_bias;
primitives[L_B_O] = (long long)lt_out_bias;
//input may have user layout (from raw image data,continuous NCHW )
// or maybe mkl layout (is previous mkl-based layer's output)
dnnLayout_t lt_in_real = (dnnLayout_t)input[MKLLayout];
if(lt_in_real==NULL) lt_in_real = lt_in_NCHW;
//create conversion and buff if necessary
dnnPrimitive_t cv_in_f = NULL; float * buf_in_f = NULL;
CHECK_ERR( try_convert(&cv_in_f, &buf_in_f, lt_in_real, lt_in_f) , err );
//create transpose if necessary
float* newPtr = NULL;
if (input[MKLLayout] == 0)
{
newPtr = (float*)malloc(inC*inH*inW*N*sizeof(float));
}
primitives[BUFFER_TRANS_INPUT] = (long long)newPtr;
//save conversion and buff
primitives[BUFFER_FORWARD_INPUT] = (long long)buf_in_f;
primitives[CONVERT_FORWARD_INPUT] = (long long)cv_in_f;
//filter layout
dnnPrimitive_t cv_filter_f = NULL; float * buf_filter_f = NULL;
CHECK_ERR( try_convert(&cv_filter_f, &buf_filter_f, lt_filter, lt_filter_f), err );
primitives[CONVERT_FORWARD_FILTER] = (long long)cv_filter_f;
primitives[BUFFER_FORWARD_FILTER] = (long long)buf_filter_f;
//save user layout for output, and create mkl buffer
//output always has mkl buffer and recorded in layer's primitive
output[CPULayout] = (long long)lt_out_CHWN;
float* buf_out_f = NULL;
CHECK_ERR( dnnAllocateBuffer_F32((void**)(&buf_out_f), lt_out_f), err );
primitives[BUFFER_FORWARD_OUTPUT] = (long long)buf_out_f;
//for bias
dnnPrimitive_t cv_bias_f = NULL; float * buf_bias_f = NULL;
dnnPrimitive_t cv_bias_b = NULL; float * buf_bias_b = NULL;
if (hasBias)
{
CHECK_ERR( try_convert(&cv_bias_f, &buf_bias_f, lt_bias_CHWN, lt_bias_f), err );
}
primitives[CONVERT_FORWARD_BIAS] = (long long)cv_bias_f;
primitives[BUFFER_FORWARD_BIAS] = (long long)buf_bias_f;
return 0;
ERR_RETURN:
return 1;
}
