Skip to content

denormals optimization for brgemm kernels #131

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
chenhu-wang wants to merge 1 commit into
base: v2.6_for_ie_master
Choose a base branch
from
Open

denormals optimization for brgemm kernels #131

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions src/cpu/x64/jit_brdgmm_dw_conv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@

#include "common/memory_tracking.hpp"
#include "common/utils.hpp"
#include <xmmintrin.h>

#include "cpu/x64/injectors/jit_uni_postops_injector.hpp"
#include "cpu/x64/jit_brdgmm_dw_conv.hpp"
Expand Down Expand Up @@ -358,6 +359,13 @@ status_t brdgmm_dw_convolution_fwd_t::execute(const exec_ctx_t &ctx) const {
const size_t dst_mb_stride = jcp.ngroups * jcp.ow * jcp.oh * jcp.dst_dsz;

parallel(jcp.nthr, [&](const int ithr, const int nthr) {
// TODO this is a hot fix for denormals, need refactor
unsigned int DENORMALS_ZERO = 0x0040;
unsigned int FLUSH_ZERO = 0x8000;
unsigned int csr = _mm_getcsr();
csr |= DENORMALS_ZERO;
csr |= FLUSH_ZERO;
_mm_setcsr(csr);
int start {0}, end {0};
balance211(work_amount, nthr, ithr, start, end);
int n {0}, chb {0}, oh {0}, owb {0};
Expand Down
13 changes: 13 additions & 0 deletions src/cpu/x64/jit_brgemm_1x1_conv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include "common/nstl.hpp"
#include "common/type_helpers.hpp"
#include "common/utils.hpp"
#include <xmmintrin.h>

#include "cpu/cpu_primitive.hpp"

Expand Down Expand Up @@ -487,6 +488,12 @@ status_t brgemm_1x1_convolution_fwd_t<isa>::execute_forward_all(
#define BRGC_WO(...) \
parallel(pd()->jcp_.nthr, [&](const int ithr, const int nthr) { \
if (ithr >= work_amount) return; \
unsigned int DENORMALS_ZERO = 0x0040; \
unsigned int FLUSH_ZERO = 0x8000; \
unsigned int csr = _mm_getcsr(); \
csr |= DENORMALS_ZERO; \
csr |= FLUSH_ZERO; \
_mm_setcsr(csr); \
brgemm_batch_element_t *const brg_batch \
= brg_batch_global + (size_t)ithr * jcp.adjusted_batch_size; \
char *const c_buffer = (jcp.use_buffer) \
Expand Down Expand Up @@ -552,6 +559,12 @@ status_t brgemm_1x1_convolution_fwd_t<isa>::execute_forward_all(
#define BRGC_WO(...) \
parallel(pd()->jcp_.nthr, [&](const int ithr, const int nthr) { \
if (ithr >= work_amount) return; \
unsigned int DENORMALS_ZERO = 0x0040; \
unsigned int FLUSH_ZERO = 0x8000; \
unsigned int csr = _mm_getcsr(); \
csr |= DENORMALS_ZERO; \
csr |= FLUSH_ZERO; \
_mm_setcsr(csr); \
brgemm_batch_element_t *const brg_batch \
= brg_batch_global + (size_t)ithr * jcp.adjusted_batch_size; \
char *const c_buffer = (jcp.use_buffer) \
Expand Down
15 changes: 15 additions & 0 deletions src/cpu/x64/jit_brgemm_conv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
#include "common/utils.hpp"
#include "cpu/cpu_primitive.hpp"
#include "cpu/x64/injectors/jit_uni_binary_injector.hpp"
#include <xmmintrin.h>

#include "cpu/x64/amx_tile_configure.hpp"
#include "cpu/x64/jit_brgemm_conv.hpp"
Expand Down Expand Up @@ -855,6 +856,13 @@ status_t brgemm_convolution_fwd_t<isa>::execute(const exec_ctx_t &ctx) const {

parallel(jcp.nthr, [&](const int ithr, const int nthr) {
if (ithr >= work_amount) return;
// TODO this is a hot fix for denormals, need refactor
unsigned int DENORMALS_ZERO = 0x0040;
unsigned int FLUSH_ZERO = 0x8000;
unsigned int csr = _mm_getcsr();
csr |= DENORMALS_ZERO;
csr |= FLUSH_ZERO;
_mm_setcsr(csr);

brgemm_batch_element_t *const __restrict brg_batch = brg_batch_global
+ static_cast<size_t>(ithr) * jcp.adjusted_batch_size;
Expand Down Expand Up @@ -1014,6 +1022,13 @@ status_t brgemm_convolution_fwd_t<isa>::cal_compensation(
const dim_t kd_b {kd_bs[k]}, kd_e {kd_es[k]}, kh_b {kh_bs[k]},
kh_e {kh_es[k]};
assert(kd_e > kd_b && kh_e > kh_b);
// TODO this is a hot fix for denormals, need refactor
unsigned int DENORMALS_ZERO = 0x0040;
unsigned int FLUSH_ZERO = 0x8000;
unsigned int csr = _mm_getcsr();
csr |= DENORMALS_ZERO;
csr |= FLUSH_ZERO;
_mm_setcsr(csr);

if (jcp.exec_type == exec_vpad && jcp.max_vpad > 0) {
const auto ow = owb * jcp.ow_block;
Expand Down
15 changes: 15 additions & 0 deletions src/cpu/x64/jit_brgemm_inner_product.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
#include "common/dnnl_thread.hpp"
#include "common/type_helpers.hpp"
#include "common/utils.hpp"
#include <xmmintrin.h>

#include "cpu/cpu_primitive.hpp"

Expand Down Expand Up @@ -297,6 +298,13 @@ status_t brgemm_inner_product_fwd_t<isa>::execute_forward(
bool ok = init_thr_groups(
ithr, nthr, nthr_ic, nthr_oc_mb, ithr_ic, ithr_oc_mb);
if (!ok) return;
// TODO this is a hot fix for denormals, need refactor
unsigned int DENORMALS_ZERO = 0x0040;
unsigned int FLUSH_ZERO = 0x8000;
unsigned int csr = _mm_getcsr();
csr |= DENORMALS_ZERO;
csr |= FLUSH_ZERO;
_mm_setcsr(csr);

int start {0}, end {0};
balance211(work_amount, nthr_oc_mb, ithr_oc_mb, start, end);
Expand Down Expand Up @@ -376,6 +384,13 @@ status_t brgemm_inner_product_fwd_t<isa>::execute_forward(
bool ok = init_thr_groups(
ithr, nthr, nthr_ic, nthr_oc_mb, ithr_ic, ithr_oc_mb);
if (!ok) return;
// TODO this is a hot fix for denormals, need refactor
unsigned int DENORMALS_ZERO = 0x0040;
unsigned int FLUSH_ZERO = 0x8000;
unsigned int csr = _mm_getcsr();
csr |= DENORMALS_ZERO;
csr |= FLUSH_ZERO;
_mm_setcsr(csr);

int ocmb_start {0}, ocmb_end {0};
int start {0}, end {0};
Expand Down