aboutsummaryrefslogtreecommitdiff
path: root/src/cpu/kernels/softmax/generic/sve2
diff options
context:
space:
mode:
Diffstat (limited to 'src/cpu/kernels/softmax/generic/sve2')
-rw-r--r--src/cpu/kernels/softmax/generic/sve2/impl.cpp289
-rw-r--r--src/cpu/kernels/softmax/generic/sve2/impl.h9
-rw-r--r--src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp12
-rw-r--r--src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp12
4 files changed, 176 insertions, 146 deletions
diff --git a/src/cpu/kernels/softmax/generic/sve2/impl.cpp b/src/cpu/kernels/softmax/generic/sve2/impl.cpp
index 8f677c62d4..98b2f5117f 100644
--- a/src/cpu/kernels/softmax/generic/sve2/impl.cpp
+++ b/src/cpu/kernels/softmax/generic/sve2/impl.cpp
@@ -23,7 +23,9 @@
*/
#include "src/cpu/kernels/softmax/generic/sve2/impl.h"
+
#include "arm_compute/core/Types.h"
+
#include "src/core/NEON/wrapper/wrapper.h"
namespace arm_compute
@@ -31,8 +33,8 @@ namespace arm_compute
namespace cpu
{
template <typename ScalarType>
-void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, float beta, bool is_log, const Window &window)
+void sve2_softmax_logits_1d_quantized(
+ const ITensor *in, const ITensor *max, void *const tmp, ITensor *out, float beta, bool is_log, const Window &window)
{
const int start_x = in->info()->valid_region().anchor.x();
const int input_width = in->info()->valid_region().shape.x();
@@ -50,162 +52,173 @@ void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, voi
const int inc_2 = static_cast<int>(2 * svcntw());
const int inc_3 = static_cast<int>(3 * svcntw());
- execute_window_loop(window, [&](const Coordinates &)
- {
- /* Get pointers */
- const auto in_ptr = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
- const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
- const auto tmp_ptr = reinterpret_cast<float *>(tmp);
+ execute_window_loop(
+ window,
+ [&](const Coordinates &)
+ {
+ /* Get pointers */
+ const auto in_ptr = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
+ const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
+ const auto tmp_ptr = reinterpret_cast<float *>(tmp);
- float sum{};
+ float sum{};
- /* Compute exponentials and sum */
- {
- /* Get max value */
- const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
- const auto vec_max = wrapper::svdup_n(max_val);
-
- /* Init sum to zero */
- auto vec_sum_0 = svdup_n_f32(0.f);
- auto vec_sum_1 = svdup_n_f32(0.f);
- auto vec_sum_2 = svdup_n_f32(0.f);
- auto vec_sum_3 = svdup_n_f32(0.f);
-
- /* Loop over row and compute exponentials and sum */
- int x = 0;
- svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
- svbool_t pg_0 = svunpklo(svunpklo(pg));
- svbool_t pg_1 = svunpkhi(svunpklo(pg));
- svbool_t pg_2 = svunpklo(svunpkhi(pg));
- svbool_t pg_3 = svunpkhi(svunpkhi(pg));
- do
+ /* Compute exponentials and sum */
{
- const auto vec_elements = svld1(pg, in_ptr + x);
- const auto vec_elements_sub = svreinterpret_u8(svsub_z(pg, vec_max, vec_elements));
+ /* Get max value */
+ const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
+ const auto vec_max = wrapper::svdup_n(max_val);
+
+ /* Init sum to zero */
+ auto vec_sum_0 = svdup_n_f32(0.f);
+ auto vec_sum_1 = svdup_n_f32(0.f);
+ auto vec_sum_2 = svdup_n_f32(0.f);
+ auto vec_sum_3 = svdup_n_f32(0.f);
+
+ /* Loop over row and compute exponentials and sum */
+ int x = 0;
+ svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+ svbool_t pg_0 = svunpklo(svunpklo(pg));
+ svbool_t pg_1 = svunpkhi(svunpklo(pg));
+ svbool_t pg_2 = svunpklo(svunpkhi(pg));
+ svbool_t pg_3 = svunpkhi(svunpkhi(pg));
+ do
+ {
+ const auto vec_elements = svld1(pg, in_ptr + x);
+ const auto vec_elements_sub = svreinterpret_u8(svsub_z(pg, vec_max, vec_elements));
+
+ auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements_sub)));
+ auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements_sub)));
+ auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements_sub)));
+ auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements_sub)));
- auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements_sub)));
- auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements_sub)));
- auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements_sub)));
- auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements_sub)));
+ if (is_log)
+ {
+ vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
+ vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
+ vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
+ vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
+ vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
+ vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
+ vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
+ vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
+ }
+ else
+ {
+ vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
+ vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
+ vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
+ vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
+ vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
+ vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
+ vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
+ vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
+ }
- if(is_log)
+ svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
+ svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
+ svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
+ svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
+
+ x += wrapper::svcnt<ScalarType>();
+ pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+ pg_0 = svunpklo(svunpklo(pg));
+ pg_1 = svunpkhi(svunpklo(pg));
+ pg_2 = svunpklo(svunpkhi(pg));
+ pg_3 = svunpkhi(svunpkhi(pg));
+ } while (svptest_any(all_true_pg, pg));
+
+ /* Reduce sum */
+ const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1),
+ svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
+ sum = svaddv_f32(all_true_pg, vec_sum);
+
+ /* Run remaining elements */
+ x = 0;
+ if (is_log)
{
- vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
- vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
- vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
- vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
- vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
- vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
- vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
- vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
+ sum = std::log(sum);
}
else
{
- vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
- vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
- vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
- vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
- vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
- vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
- vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
- vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
+ sum = 256.f / sum;
}
-
- svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
- svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
- svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
- svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
-
- x += wrapper::svcnt<ScalarType>();
- pg = wrapper::svwhilelt<ScalarType>(x, input_width);
- pg_0 = svunpklo(svunpklo(pg));
- pg_1 = svunpkhi(svunpklo(pg));
- pg_2 = svunpklo(svunpkhi(pg));
- pg_3 = svunpkhi(svunpkhi(pg));
}
- while(svptest_any(all_true_pg, pg));
- /* Reduce sum */
- const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1), svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
- sum = svaddv_f32(all_true_pg, vec_sum);
-
- /* Run remaining elements */
- x = 0;
- if(is_log)
- {
- sum = std::log(sum);
- }
- else
+ /* Normalize exponentials */
{
- sum = 256.f / sum;
- }
- }
-
- /* Normalize exponentials */
- {
- constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
- /* Loop over row and compute softmax */
- int x = 0;
- svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
- svbool_t pg_0 = svunpklo(svunpklo(pg));
- svbool_t pg_1 = svunpkhi(svunpklo(pg));
- svbool_t pg_2 = svunpklo(svunpkhi(pg));
- svbool_t pg_3 = svunpkhi(svunpkhi(pg));
- do
- {
- auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
- auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
- auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
- auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
-
- svfloat32_t res_0{};
- svfloat32_t res_1{};
- svfloat32_t res_2{};
- svfloat32_t res_3{};
-
- if(is_log)
+ constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
+ /* Loop over row and compute softmax */
+ int x = 0;
+ svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+ svbool_t pg_0 = svunpklo(svunpklo(pg));
+ svbool_t pg_1 = svunpkhi(svunpklo(pg));
+ svbool_t pg_2 = svunpklo(svunpkhi(pg));
+ svbool_t pg_3 = svunpkhi(svunpkhi(pg));
+ do
{
- res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
- res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
- res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
- res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
- }
- else
- {
- res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
- res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
- res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
- res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
+ auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
+ auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
+ auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
+ auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
+
+ svfloat32_t res_0{};
+ svfloat32_t res_1{};
+ svfloat32_t res_2{};
+ svfloat32_t res_3{};
- if(is_qasymm8_signed)
+ if (is_log)
{
- const auto offset_vec = svdup_n_f32(128.f);
- res_0 = svsub_z(pg_0, res_0, offset_vec);
- res_1 = svsub_z(pg_1, res_1, offset_vec);
- res_2 = svsub_z(pg_2, res_2, offset_vec);
- res_3 = svsub_z(pg_3, res_3, offset_vec);
+ res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
+ res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
+ res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
+ res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
+ }
+ else
+ {
+ res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
+ res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
+ res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
+ res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
+
+ if (is_qasymm8_signed)
+ {
+ const auto offset_vec = svdup_n_f32(128.f);
+ res_0 = svsub_z(pg_0, res_0, offset_vec);
+ res_1 = svsub_z(pg_1, res_1, offset_vec);
+ res_2 = svsub_z(pg_2, res_2, offset_vec);
+ res_3 = svsub_z(pg_3, res_3, offset_vec);
+ }
}
- }
- // Store value
- const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
- svst1(pg, out_ptr + x, out);
- x += wrapper::svcnt<ScalarType>();
- pg = wrapper::svwhilelt<ScalarType>(x, input_width);
- pg_0 = svunpklo(svunpklo(pg));
- pg_1 = svunpkhi(svunpklo(pg));
- pg_2 = svunpklo(svunpkhi(pg));
- pg_3 = svunpkhi(svunpkhi(pg));
+ // Store value
+ const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
+ svst1(pg, out_ptr + x, out);
+ x += wrapper::svcnt<ScalarType>();
+ pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+ pg_0 = svunpklo(svunpklo(pg));
+ pg_1 = svunpkhi(svunpklo(pg));
+ pg_2 = svunpklo(svunpkhi(pg));
+ pg_3 = svunpkhi(svunpkhi(pg));
+ } while (svptest_any(all_true_pg, pg));
}
- while(svptest_any(all_true_pg, pg));
- }
- },
- in_it, max_it, out_it);
+ },
+ in_it, max_it, out_it);
}
-template void sve2_softmax_logits_1d_quantized<qasymm8_signed_t>(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, float beta, bool is_log, const Window &window);
-template void sve2_softmax_logits_1d_quantized<qasymm8_t>(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, float beta, bool is_log, const Window &window);
+template void sve2_softmax_logits_1d_quantized<qasymm8_signed_t>(const ITensor *in,
+ const ITensor *max,
+ void *const tmp,
+ ITensor *out,
+ float beta,
+ bool is_log,
+ const Window &window);
+template void sve2_softmax_logits_1d_quantized<qasymm8_t>(const ITensor *in,
+ const ITensor *max,
+ void *const tmp,
+ ITensor *out,
+ float beta,
+ bool is_log,
+ const Window &window);
} // namespace cpu
} // namespace arm_compute
diff --git a/src/cpu/kernels/softmax/generic/sve2/impl.h b/src/cpu/kernels/softmax/generic/sve2/impl.h
index abbcc15181..33fcc26cda 100644
--- a/src/cpu/kernels/softmax/generic/sve2/impl.h
+++ b/src/cpu/kernels/softmax/generic/sve2/impl.h
@@ -31,8 +31,13 @@ namespace arm_compute
namespace cpu
{
template <typename ScalarType>
-void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, float beta, bool is_log, const Window &window);
+void sve2_softmax_logits_1d_quantized(const ITensor *in,
+ const ITensor *max,
+ void *const tmp,
+ ITensor *out,
+ float beta,
+ bool is_log,
+ const Window &window);
} // namespace cpu
} // namespace arm_compute
#endif /* SRC_CORE_SVE2_KERNELS_SOFTMAX_IMPL_H */
diff --git a/src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp b/src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp
index 810035eb9c..95623786b3 100644
--- a/src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp
+++ b/src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp
@@ -23,16 +23,22 @@
*/
#include "arm_compute/core/Helpers.h"
+
#include "src/cpu/kernels/softmax/generic/sve2/impl.h"
namespace arm_compute
{
namespace cpu
{
-void sve2_qasymm8_softmax(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, const float beta, bool is_log, const Window &window)
+void sve2_qasymm8_softmax(const ITensor *in,
+ const ITensor *max,
+ void *const tmp,
+ ITensor *out,
+ const float beta,
+ bool is_log,
+ const Window &window)
{
return sve2_softmax_logits_1d_quantized<qasymm8_t>(in, max, tmp, out, beta, is_log, window);
}
-}
+} // namespace cpu
} // namespace arm_compute
diff --git a/src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp b/src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp
index 283b55e9ce..c20462fcef 100644
--- a/src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp
+++ b/src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp
@@ -23,16 +23,22 @@
*/
#include "arm_compute/core/Helpers.h"
+
#include "src/cpu/kernels/softmax/generic/sve2/impl.h"
namespace arm_compute
{
namespace cpu
{
-void sve2_qasymm8_signed_softmax(const ITensor *in, const ITensor *max, void *const tmp,
- ITensor *out, const float beta, bool is_log, const Window &window)
+void sve2_qasymm8_signed_softmax(const ITensor *in,
+ const ITensor *max,
+ void *const tmp,
+ ITensor *out,
+ const float beta,
+ bool is_log,
+ const Window &window)
{
return sve2_softmax_logits_1d_quantized<qasymm8_signed_t>(in, max, tmp, out, beta, is_log, window);
}
-}
+} // namespace cpu
} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-12 00:28:21 +0000