typename data_t_acc, typename data_t_dst>
void compute_ref_conv_fwd(const mkldnn_convolution_desc_t &conv_desc,
const memory &src, const memory &weights, const memory &bias, const memory &dst,
- bool with_relu, float eltwise_alpha)
+ bool with_relu, float eltwise_alpha, const float* depthwise_weights)
{
int MB = conv_desc.src_desc.dims[0];
int G = conv_desc.weights_desc.ndims == 5 ? conv_desc.weights_desc.dims[0] : 1;
data_t_src *src_data = (data_t_src *)src.get_data_handle();
data_t_wei *weights_data = (data_t_wei *)weights.get_data_handle();
- data_t_dst *bias_data = (data_t_dst *)bias.get_data_handle();
+ float *bias_data = (float *)bias.get_data_handle();
data_t_dst *dst_data = (data_t_dst *)dst.get_data_handle();
const memory::desc src_d = src.get_primitive_desc().desc();
a += src_data[map_index(src_d, iidx)]
* weights_data[map_index(
weights_d, widx)];
-
-
}
}
}
a_fp += bias_data[G > 1 ? g : oc];
+ if (depthwise_weights)
+ a_fp *= depthwise_weights[G > 1 ? g : oc];
+
if (with_relu) {
a_fp = (a_fp > 0) ? a_fp : eltwise_alpha * a_fp;
}
- dst_data[map_index(dst_d, oidx)] = (data_t_dst) a_fp;
+ using D = memory::data_type;
+ if (data_traits<data_t_dst>::data_type != D::f32){
+ a_fp = nearbyintf(a_fp);
+ }
+
+ dst_data[map_index(dst_d, oidx)] = (data_t_dst)a_fp;
}
);
}
memory::data_type data_type_src = data_traits<data_t_src>::data_type;
memory::data_type data_type_dst = data_traits<data_t_dst>::data_type;
memory::data_type data_type_wei = data_traits<data_t_wei>::data_type;
- memory::data_type data_type_bia = data_traits<data_t_wei>::data_type;
+ memory::data_type data_type_bia = data_traits<float>::data_type;
+
+ bool is_int8 = data_type_src == mkldnn_u8 || data_type_src == mkldnn_s8;
test_convolution_dw_conv_sizes_t cd = p.sizes;
int conv2_oh = (conv1_oh - ((cd.conv2_kh - 1) + 1) + 2 * cd.conv2_padh) / cd.conv2_strh + 1;
int conv2_ow = (conv1_ow - ((cd.conv2_kw - 1) + 1) + 2 * cd.conv2_padw) / cd.conv2_strw + 1;
- std::vector<int> conv1_padR = { cd.conv1_padh, cd.conv1_padw };
+ std::vector<ptrdiff_t> conv1_padR = { cd.conv1_padh, cd.conv1_padw };
conv1_padR[0] += conv2_oh - conv1_oh;
conv1_padR[1] += conv2_ow - conv1_ow;
auto conv2_dst = memory({conv2_dst_desc, eng});
fill_data<data_t_src>(conv1_src.get_primitive_desc().get_size()
- / sizeof(data_t_src), (data_t_src *)conv1_src.get_data_handle(), 1., true);
+ / sizeof(data_t_src), (data_t_src *)conv1_src.get_data_handle(), (data_t_src)1, (data_t_src)1);
fill_data<data_t_wei>(
conv1_weights.get_primitive_desc().get_size()
- / sizeof(data_t_wei),(data_t_wei *)conv1_weights.get_data_handle(), 1., true);
- fill_data<data_t_wei>(
+ / sizeof(data_t_wei),(data_t_wei *)conv1_weights.get_data_handle(), (data_t_wei)1, (data_t_wei)1);
+ fill_data<float>(
conv1_bias.get_primitive_desc().get_size()
- / sizeof(data_t_wei),(data_t_wei *)conv1_bias.get_data_handle(), 1., true);
+ / sizeof(float),(float *)conv1_bias.get_data_handle(), 1., true);
fill_data<data_t_wei>(
conv2_weights.get_primitive_desc().get_size()
- / sizeof(data_t_wei),(data_t_wei *)conv2_weights.get_data_handle(), 1., true);
- fill_data<data_t_wei>(
+ / sizeof(data_t_wei),(data_t_wei *)conv2_weights.get_data_handle(), (data_t_wei)1, (data_t_wei)1);
+ fill_data<float>(
conv2_bias.get_primitive_desc().get_size()
- / sizeof(data_t_wei),(data_t_wei *)conv2_bias.get_data_handle(), 1., true);
+ / sizeof(float),(float *)conv2_bias.get_data_handle(), 1., true);
+
+// auto conv1_depthwise_weights_desc = create_md({ cd.conv2_oc }, mkldnn::memory::data_type::f32, memory::x);
+// auto conv1_depthwise_weights = memory({conv1_depthwise_weights_desc, eng});
+ std::vector<float> conv1_depthwise_weights;
+ conv1_depthwise_weights.resize(cd.conv1_oc);
+ fill_data<float>(conv1_depthwise_weights.size(), &conv1_depthwise_weights[0], 1.f / ((float)cd.ic), 1.f / ((float)cd.ic * cd.conv1_kh * cd.conv1_kw));
+
+ std::vector<float> conv2_depthwise_weights;
+ conv2_depthwise_weights.resize(cd.conv1_oc);
+ fill_data<float>(conv2_depthwise_weights.size(), &conv2_depthwise_weights[0], 1.f / ((float)cd.conv2_oc), 1.f / ((float)cd.conv2_oc * cd.conv2_kh * cd.conv2_kw));
+
+ std::vector<float> conv2_depthwise_bias;
+ conv2_depthwise_bias.resize(cd.conv1_oc);
+// fill_data<float>(conv2_depthwise_bias.size(), &conv2_depthwise_bias[0], 1., true);
+ memset(&conv2_depthwise_bias[0], 0, conv2_depthwise_bias.size() * sizeof(float));
+
+// auto conv2_depthwise_weights_desc = create_md({ cd.conv2_oc }, mkldnn::memory::data_type::f32, memory::x);
+// auto conv2_depthwise_bias_desc = create_md({ cd.conv2_oc }, mkldnn::memory::data_type::f32, memory::x);
+//
+// auto conv2_depthwise_weights = memory({conv2_depthwise_weights_desc, eng});
+// auto conv2_depthwise_bias = memory({conv2_depthwise_bias_desc, eng});
+
+// fill_data<float>(conv2_depthwise_weights.get_primitive_desc().get_size() / sizeof(float),
+// (float *)conv2_depthwise_weights.get_data_handle(), 1., true);
+// memset((float*)conv2_depthwise_bias.get_data_handle(), 0, conv2_depthwise_bias.get_primitive_desc().get_size());
mkldnn::post_ops conv1_post_ops;
conv1_post_ops.append_eltwise(1.0, mkldnn::algorithm::eltwise_relu, 0.0f, 0.0f);
conv1_post_ops.append_dw_conv(conv1_oh, conv1_ow, cd.conv2_kh, cd.conv2_kw, cd.conv2_strh, cd.conv2_strw,
static_cast<const float*>(conv2_weights.get_data_handle()),
static_cast<const float*>(conv2_bias.get_data_handle()));
+
+ if (is_int8)
+ conv1_post_ops.append_depthwise(depthwise_scale_shift, &conv2_depthwise_weights[0], &conv2_depthwise_bias[0]);
+
conv1_post_ops.append_eltwise(1.0, mkldnn::algorithm::eltwise_relu, 0.0f, 0.0f);
mkldnn::primitive_attr conv1_attr;
+
+ if (is_int8) {
+ conv1_attr.set_int_output_round_mode(mkldnn::round_nearest);
+ conv1_attr.set_output_scales(1 << 1 /*through C dim*/, conv1_depthwise_weights);
+ }
+
conv1_attr.set_post_ops(conv1_post_ops);
auto conv1_primitive_desc = convolution_forward::primitive_desc(conv1_desc, conv1_attr, eng);
auto conv1_dst_ref = memory({conv1_dst_desc_ref, eng});
auto conv2_dst_ref = memory({conv2_dst_desc, eng});
- compute_ref_conv_fwd<data_t_src, data_t_wei, data_t_acc, data_t_dst>(conv1_desc_ref.data, conv1_src, conv1_weights, conv1_bias, conv1_dst_ref, true, 0.0f);
- compute_ref_conv_fwd<data_t_dst, data_t_wei, data_t_acc, data_t_dst>(conv2_desc.data, conv1_dst_ref, conv2_weights, conv2_bias, conv2_dst_ref, true, 0.0f);
+
+ auto conv1_depthwise_weights_data = is_int8 ? &conv1_depthwise_weights[0] : nullptr;
+ auto conv2_depthwise_weights_data = is_int8 ? &conv2_depthwise_weights[0] : nullptr;
+
+ compute_ref_conv_fwd<data_t_src, data_t_wei, data_t_acc, data_t_dst>(conv1_desc_ref.data,
+ conv1_src, conv1_weights, conv1_bias, conv1_dst_ref, true, 0.0f, conv1_depthwise_weights_data);
+ compute_ref_conv_fwd<data_t_dst, data_t_wei, data_t_acc, data_t_dst>(conv2_desc.data,
+ conv1_dst_ref, conv2_weights, conv2_bias, conv2_dst_ref, true, 0.0f, conv2_depthwise_weights_data);
compare_data<data_t_dst>(conv2_dst_ref, conv2_dst);
}
projects / platform / upstream / dldt.git / blobdiff