frozenleaves/oneDNN
1
0
Fork 0
mirror of https://github.com/uxlfoundation/oneDNN.git synced 2025-10-20 10:03:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

benchdnn: postops: skip select postop cases for gpu device

Browse Source
This commit is contained in:
Ankit Manerikar
2025-05-26 07:54:52 -07:00
parent df7d742a09
commit d89d6469b0
22 changed files with 35 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/benchdnn/binary/binary.cpp
View File

@ -128,6 +128,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
std::vector<dnnl_data_type_t> dts = {prb->sdt[0], prb->sdt[1], prb->ddt};
skip_unimplemented_data_type(dts, prb->dir, res);
skip_unimplemented_arg_scale(prb->attr, res);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_binary);
if (is_gpu()) {

1
tests/benchdnn/bnorm/bnorm.cpp
View File

@ -455,6 +455,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->dt}, prb->dir, res);
skip_unimplemented_sum_po(
prb->attr, res, dnnl_batch_normalization, prb->dt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_batch_normalization);
// Non-zero alpha is not supported for training in general.

1
tests/benchdnn/brgemm/brgemm.cpp
View File

@ -495,6 +495,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
prb->dir, res);
skip_unimplemented_sum_po(
prb->attr, res, dnnl_gemm, prb->src_dt(), prb->dst_dt());
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_gemm);
// Unconditionally skip remaining unimplemented cases.

1
tests/benchdnn/concat/concat.cpp
View File

@ -116,6 +116,7 @@ int fill_src(int input_idx, dnnl_data_type_t dt, dnn_mem_t &mem_dt,
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->sdt, prb->ddt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_concat, prb->sdt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_concat);
skip_unimplemented_arg_scale(prb->attr, res);

1
tests/benchdnn/conv/conv.cpp
View File

@ -378,6 +378,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_convolution,
prb->get_dt(SRC), prb->get_dt(DST));
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_convolution);
if (is_cpu()) {

1
tests/benchdnn/deconv/deconv.cpp
View File

@ -349,6 +349,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
prb->dir, res);
skip_unimplemented_sum_po(
prb->attr, res, dnnl_deconvolution, prb->get_dt(SRC));
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_deconvolution);
// GPU supports only post ops and all but x8s8bf16 and f32xf16f32 cfg

14
tests/benchdnn/dnnl_common.cpp
View File

@ -803,6 +803,20 @@ void skip_unimplemented_sum_po(const attr_t &attr, res_t *res,
}
}
void skip_unimplemented_binary_po(const attr_t &attr, res_t *res) {
const auto &po = attr.post_ops;
if (po.is_def()) return;
if (is_gpu()) {
const int select_idx = po.find(attr_t::post_ops_t::kind_t::SELECT);
if (select_idx != -1) {
res->state = SKIPPED;
res->reason = skip_reason::case_not_supported;
return;
}
}
}
void skip_unimplemented_prelu_po(
const attr_t &attr, res_t *res, dnnl_primitive_kind_t pkind) {
const auto &po = attr.post_ops;

1
tests/benchdnn/dnnl_common.hpp
View File

@ -262,6 +262,7 @@ void skip_unimplemented_data_type(
void skip_unimplemented_sum_po(const attr_t &attr, res_t *res,
dnnl_primitive_kind_t pkind, dnnl_data_type_t src_dt,
dnnl_data_type_t dst_dt = dnnl_data_type_undef);
void skip_unimplemented_binary_po(const attr_t &attr, res_t *res);
void skip_unimplemented_prelu_po(
const attr_t &attr, res_t *res, dnnl_primitive_kind_t pkind);
void skip_invalid_inplace(res_t *res, dnnl_data_type_t sdt,

1
tests/benchdnn/eltwise/eltwise.cpp
View File

@ -287,6 +287,7 @@ int fill_data(const prb_t *prb, data_kind_t kind, dnn_mem_t &mem_dt,
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->dt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_eltwise, prb->dt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_eltwise);
if (is_gpu() && (prb->dt == dnnl_f8_e5m2 || prb->dt == dnnl_f8_e4m3)

1
tests/benchdnn/ip/ip.cpp
View File

@ -246,6 +246,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_inner_product,
prb->get_dt(SRC), prb->get_dt(DST));
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_inner_product);
if (is_cpu()) {

1
tests/benchdnn/lnorm/lnorm.cpp
View File

@ -455,6 +455,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
{prb->dt[0], prb->dt[1], prb->ss_dt}, prb->dir, res);
skip_unimplemented_sum_po(
prb->attr, res, dnnl_layer_normalization, prb->dt[0]);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_layer_normalization);
if (is_gpu() && prb->attr.post_ops.len() != 0) {

1
tests/benchdnn/lrn/lrn.cpp
View File

@ -118,6 +118,7 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->dt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_lrn, prb->dt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_lrn);
}

1
tests/benchdnn/matmul/matmul.cpp
View File

@ -485,6 +485,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
prb->dir, res);
skip_unimplemented_sum_po(
prb->attr, res, dnnl_matmul, prb->src_dt(), prb->dst_dt());
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_matmul);
if ((is_nvidia_gpu() || is_amd_gpu()) && !prb->sparse_options.is_def()) {

1
tests/benchdnn/pool/pool.cpp
View File

@ -144,6 +144,7 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->src_dt(), prb->dst_dt()}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_pooling, prb->src_dt());
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_pooling);
if (is_cpu() && prb->src_dt() != prb->dst_dt()) {

1
tests/benchdnn/prelu/prelu.cpp
View File

@ -142,6 +142,7 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type(prb->sdt, FWD_D, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_prelu, prb->sdt[0]);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_prelu);
}

1
tests/benchdnn/reduction/reduction.cpp
View File

@ -204,6 +204,7 @@ int fill_dst(const prb_t *prb, dnn_mem_t &mem_dt, dnn_mem_t &mem_fp) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->sdt, prb->ddt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_reduction, prb->sdt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_reduction);
}

1
tests/benchdnn/reorder/reorder.cpp
View File

@ -208,6 +208,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
const auto ddt = prb->ddt;
skip_unimplemented_data_type({sdt, ddt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_reorder, sdt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_reorder);
const bool s32_src_ok = IMPLICATION(sdt == dnnl_s32,

1
tests/benchdnn/resampling/resampling.cpp
View File

@ -111,6 +111,7 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->sdt, prb->ddt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_resampling, prb->sdt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_resampling);
}

1
tests/benchdnn/rnn/rnn.cpp
View File

@ -778,6 +778,7 @@ void skip_unimplemented_prb(const prb_t *prb_, res_t *res) {
dir_t dir = str2dir(prop2str(prb.prop));
skip_unimplemented_data_type({prb.cfg[SRC_LAYER].dt}, dir, res);
skip_unimplemented_sum_po(prb.attr, res, dnnl_rnn, prb.cfg[SRC_LAYER].dt);
skip_unimplemented_binary_po(prb.attr, res);
skip_unimplemented_prelu_po(prb.attr, res, dnnl_rnn);
if (is_cpu()) {

1
tests/benchdnn/shuffle/shuffle.cpp
View File

@ -102,6 +102,7 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->dt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_shuffle, prb->dt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_shuffle);
}

1
tests/benchdnn/softmax/softmax.cpp
View File

@ -222,6 +222,7 @@ int fill_data_bwd(data_kind_t data_kind, const prb_t *prb, dnn_mem_t &mem_dt,
void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
skip_unimplemented_data_type({prb->sdt, prb->ddt}, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_softmax, prb->sdt);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_softmax);
if (prb->attr.post_ops.find(attr_t::post_ops_t::kind_t::SUM) != -1) {

1
tests/benchdnn/sum/sum.cpp
View File

@ -99,6 +99,7 @@ void skip_unimplemented_prb(const prb_t *prb, res_t *res) {
dts.push_back(prb->ddt);
skip_unimplemented_data_type(dts, prb->dir, res);
skip_unimplemented_sum_po(prb->attr, res, dnnl_sum, prb->sdt[0]);
skip_unimplemented_binary_po(prb->attr, res);
skip_unimplemented_prelu_po(prb->attr, res, dnnl_sum);
}