Skip to content

Commit 6f5eb48

Browse files
yury-intelyury-intel
committed
save
1 parent ece1d37 commit 6f5eb48

File tree

1 file changed

+41
-28
lines changed

1 file changed

+41
-28
lines changed

‎inference-engine/src/mkldnn_plugin/nodes/mkldnn_def_conv_node.cpp‎

Lines changed: 41 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -80,6 +80,7 @@ struct jit_uni_def_conv_kernel_f32 : public jit_uni_def_conv_kernel, public jit_
8080
reg64_t reg_ow_pos = rdx;
8181
reg64_t aux_reg_output = reg_ow_pos;
8282
reg64_t reg_dg_iter = reg_output;
83+
reg64_t reg_gr_iter = rsp;
8384
reg64_t aux_reg_input = rax;
8485
reg64_t aux2_reg_input = reg_kernel;
8586
reg64_t reg_ic_iter = rbx;
@@ -853,7 +854,6 @@ struct jit_uni_def_conv_kernel_f32 : public jit_uni_def_conv_kernel, public jit_
853854
L(oc_unrolled_loop); {
854855
cmp(reg_oc_work, jcp_.nb_oc_blocking * jcp_.oc_block);
855856
jl(oc_main_loop, T_NEAR);
856-
857857
ic_loop(ow_step, jcp_.nb_oc_blocking, jcp_.oc_block);
858858
store_output(ow_step, jcp_.nb_oc_blocking, jcp_.oc_block);
859859

@@ -869,7 +869,6 @@ struct jit_uni_def_conv_kernel_f32 : public jit_uni_def_conv_kernel, public jit_
869869
L(oc_main_loop); {
870870
cmp(reg_oc_work, jcp_.oc_block);
871871
jl(oc_tail, T_NEAR);
872-
873872
ic_loop(ow_step, 1, jcp_.oc_block);
874873
store_output(ow_step, 1, jcp_.oc_block);
875874

@@ -987,17 +986,18 @@ void MKLDNNDeformableConvolutionNode::initSupportedPrimitiveDescriptors() {
987986
config.outConfs[0].inPlace = -1;
988987

989988
impl_desc_type impl_type;
990-
if (mayiuse(cpu::x64::avx512_common)) {
991-
impl_type = impl_desc_type::jit_avx512;
992-
} else if (mayiuse(cpu::x64::avx2)) {
993-
impl_type = impl_desc_type::jit_avx2;
994-
} else if (mayiuse(cpu::x64::sse41)) {
995-
impl_type = impl_desc_type::jit_sse42;
996-
} else {
997-
impl_type = impl_desc_type::ref;
998-
}
999-
1000-
if (mayiuse(cpu::x64::sse41)) {
989+
// if (mayiuse(cpu::x64::avx512_common)) {
990+
// impl_type = impl_desc_type::jit_avx512;
991+
// } else if (mayiuse(cpu::x64::avx2)) {
992+
// impl_type = impl_desc_type::jit_avx2;
993+
// } else if (mayiuse(cpu::x64::sse41)) {
994+
// impl_type = impl_desc_type::jit_sse42;
995+
// } else {
996+
// impl_type = impl_desc_type::ref;
997+
// }
998+
impl_type = impl_desc_type::ref;
999+
1000+
if (false && mayiuse(cpu::x64::sse41)) {
10011001
// optimzed implementation
10021002
auto dataFormat = memory::format_tag::nhwc;
10031003
auto offFormat = memory::format_tag::nchw;
@@ -1062,9 +1062,9 @@ void MKLDNNDeformableConvolutionNode::createPrimitive() {
10621062
jcp.oh = dstDims[2];
10631063
jcp.ow = dstDims[3];
10641064

1065-
bool with_groups = group > 1;
1066-
jcp.kh = weiDims[with_groups + 2];
1067-
jcp.kw = weiDims[with_groups + 3];
1065+
// bool with_groups = group > 1;
1066+
jcp.kh = weiDims[2];
1067+
jcp.kw = weiDims[3];
10681068

10691069
jcp.t_pad = paddingL[0];
10701070
jcp.l_pad = paddingL[1];
@@ -1097,13 +1097,13 @@ void MKLDNNDeformableConvolutionNode::createPrimitive() {
10971097

10981098
jcp.nthr = dnnl_get_max_threads();
10991099

1100-
if (mayiuse(cpu::x64::avx512_common)) {
1101-
def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::avx512_common>(jcp));
1102-
} else if (mayiuse(cpu::x64::avx2)) {
1103-
def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::avx2>(jcp));
1104-
} else if (mayiuse(cpu::x64::sse41)) {
1105-
def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::sse41>(jcp));
1106-
}
1100+
// if (mayiuse(cpu::x64::avx512_common)) {
1101+
// def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::avx512_common>(jcp));
1102+
// } else if (mayiuse(cpu::x64::avx2)) {
1103+
// def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::avx2>(jcp));
1104+
// } else if (mayiuse(cpu::x64::sse41)) {
1105+
// def_conv_kernel.reset(new jit_uni_def_conv_kernel_f32<cpu::x64::sse41>(jcp));
1106+
// }
11071107

11081108
if (def_conv_kernel)
11091109
def_conv_kernel->create_ker();
@@ -1147,7 +1147,7 @@ void MKLDNNDeformableConvolutionNode::executeReference(const float* src, const f
11471147

11481148
for (int ic = 0; ic < IC; ic++) {
11491149
const float *data_im_ptr = src + mb * src_strides[0] + (g * IC + ic) * src_strides[1] + h_in * src_strides[2] + w_in * src_strides[3];
1150-
const int deformable_group_index = ic / channel_per_deformable_group;
1150+
const int deformable_group_index = (IC * g + ic) / channel_per_deformable_group;
11511151
const float *data_offset_ptr = offsets + mb * off_strides[0] + (deformable_group_index * 2 * KH * KW) * off_strides[1];
11521152
const float *modulation_offset_ptr = nullptr;
11531153
if (modulation != nullptr) {
@@ -1165,7 +1165,20 @@ void MKLDNNDeformableConvolutionNode::executeReference(const float* src, const f
11651165

11661166
const float h_im = h_in + map_h; // absolute pixel index with offset
11671167
const float w_im = w_in + map_w; // absolute pixel index with offset
1168-
if (h_im >= 0 && w_im >= 0 && h_im < IH && w_im < IW) {
1168+
bool skip_compute;
1169+
if (with_bilinear_pad) {
1170+
skip_compute = !(static_cast<int>(w_im) > -1 &&
1171+
static_cast<int>(w_im) < IW &&
1172+
static_cast<int>(h_im) > -1 &&
1173+
static_cast<int>(h_im) < IH);
1174+
} else {
1175+
skip_compute = !(w_im >= 0 &&
1176+
w_im < IW &&
1177+
h_im >= 0 &&
1178+
h_im < IH);
1179+
}
1180+
if (!skip_compute) {
1181+
// if (h_im >= 0 && w_im >= 0 && h_im < IH && w_im < IW) {
11691182
const int cur_height = IH - h_in;
11701183
const int cur_width = IW - w_in;
11711184
int h_low = std::max(static_cast<int>(floorf(map_h)), 0);
@@ -1192,8 +1205,8 @@ void MKLDNNDeformableConvolutionNode::executeReference(const float* src, const f
11921205
modulation_scalar = modulation_offset_ptr[modulation_index];
11931206
}
11941207

1195-
const float weight = with_groups ? weights[g * wei_strides[0] + oc * wei_strides[1] + ic * wei_strides[2] + kh * wei_strides[3] +
1196-
kw * wei_strides[4]]
1208+
const float weight = with_groups ? weights[(g + oc / G) * wei_strides[0] + ic * wei_strides[1] + kh * wei_strides[2] +
1209+
kw * wei_strides[3]]
11971210
: weights[oc * wei_strides[0] + ic * wei_strides[1] + kh * wei_strides[2] + kw * wei_strides[3]];
11981211
d += val * weight * modulation_scalar;
11991212
}
@@ -1205,7 +1218,7 @@ void MKLDNNDeformableConvolutionNode::executeReference(const float* src, const f
12051218
};
12061219

12071220
parallel_nd(G, MB, OC, OH, OW,
1208-
[&](int g, int mb, int oc, int oh, int ow) {
1221+
[&](int g, int mb, int oc, int oh, int ow) {
12091222
dst[mb * dst_strides[0] + (g * OC + oc) * dst_strides[1] + oh * dst_strides[2] + ow * dst_strides[3]] = ker(g, mb, oc, oh, ow);
12101223
});
12111224
}

0 commit comments

Comments
 (0)