diff options
Diffstat (limited to 'src/cpu/kernels/softmax/generic/neon/impl.cpp')
| -rw-r--r-- | src/cpu/kernels/softmax/generic/neon/impl.cpp | 281 |
1 files changed, 146 insertions, 135 deletions
diff --git a/src/cpu/kernels/softmax/generic/neon/impl.cpp b/src/cpu/kernels/softmax/generic/neon/impl.cpp index f07fd2fb27..5d6e6a4f80 100644 --- a/src/cpu/kernels/softmax/generic/neon/impl.cpp +++ b/src/cpu/kernels/softmax/generic/neon/impl.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "src/cpu/kernels/softmax/generic/neon/impl.h" + #include "support/SaturateCast.h" namespace arm_compute @@ -32,11 +33,10 @@ template void neon_logits_1d_max<qasymm8_signed_t>(const ITensor *in, ITensor *o template void neon_logits_1d_max<qasymm8_t>(const ITensor *in, ITensor *out, const Window &window); template <typename T> -void neon_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp, - ITensor *out, float beta, bool is_log, const Window &window) +void neon_softmax_logits_1d_quantized( + const ITensor *in, const ITensor *max, void *const tmp, ITensor *out, float beta, bool is_log, const Window &window) { - static_assert(std::is_same<T, qasymm8_t>::value - || std::is_same<T, qasymm8_signed_t>::value, + static_assert(std::is_same<T, qasymm8_t>::value || std::is_same<T, qasymm8_signed_t>::value, "quantized type should be either qasymm8_t or qasymm8_signed_t."); const int start_x = in->info()->valid_region().anchor.x(); @@ -50,163 +50,174 @@ void neon_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, voi Iterator out_it(out, window); constexpr int vec_size = 16; - execute_window_loop(window, [&](const Coordinates &) - { - /* Get pointers */ - const auto in_ptr = reinterpret_cast<const T *>(in_it.ptr()) + start_x; - const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x; - const auto tmp_ptr = reinterpret_cast<float *>(tmp); - - float sum{}; - float sum_inversed{}; - - /* Compute exponentials and sum */ + execute_window_loop( + window, + [&](const Coordinates &) { - /* Get max value */ - const auto max_val = *reinterpret_cast<const T *>(max_it.ptr()); - const auto vec_max = wrapper::vdup_n(max_val, wrapper::traits::vector_128_tag{}); + /* Get pointers */ + const auto in_ptr = reinterpret_cast<const T *>(in_it.ptr()) + start_x; + const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x; + const auto tmp_ptr = reinterpret_cast<float *>(tmp); - /* Init sum to zero */ - float32x4x4_t vec_sum = - { - vdupq_n_f32(0.f), - vdupq_n_f32(0.f), - vdupq_n_f32(0.f), - vdupq_n_f32(0.f), - }; - - /* Loop over row and compute exponentials and sum */ - int x = 0; - for(; x <= (input_width - vec_size); x += vec_size) - { - auto vec_elements = wrapper::vloadq(in_ptr + x); - vec_elements = wrapper::vqsub(vec_max, vec_elements); - auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements); + float sum{}; + float sum_inversed{}; - if(is_log) - { - vec_elements_flt.val[0] = vmulq_f32(vec_elements_flt.val[0], scale_beta_vec); - vec_elements_flt.val[1] = vmulq_f32(vec_elements_flt.val[1], scale_beta_vec); - vec_elements_flt.val[2] = vmulq_f32(vec_elements_flt.val[2], scale_beta_vec); - vec_elements_flt.val[3] = vmulq_f32(vec_elements_flt.val[3], scale_beta_vec); - vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vexpq_f32(vec_elements_flt.val[0])); - vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vexpq_f32(vec_elements_flt.val[1])); - vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vexpq_f32(vec_elements_flt.val[2])); - vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vexpq_f32(vec_elements_flt.val[3])); - } - else + /* Compute exponentials and sum */ + { + /* Get max value */ + const auto max_val = *reinterpret_cast<const T *>(max_it.ptr()); + const auto vec_max = wrapper::vdup_n(max_val, wrapper::traits::vector_128_tag{}); + + /* Init sum to zero */ + float32x4x4_t vec_sum = { + vdupq_n_f32(0.f), + vdupq_n_f32(0.f), + vdupq_n_f32(0.f), + vdupq_n_f32(0.f), + }; + + /* Loop over row and compute exponentials and sum */ + int x = 0; + for (; x <= (input_width - vec_size); x += vec_size) { - vec_elements_flt.val[0] = vexpq_f32(vmulq_f32(vec_elements_flt.val[0], scale_beta_vec)); - vec_elements_flt.val[1] = vexpq_f32(vmulq_f32(vec_elements_flt.val[1], scale_beta_vec)); - vec_elements_flt.val[2] = vexpq_f32(vmulq_f32(vec_elements_flt.val[2], scale_beta_vec)); - vec_elements_flt.val[3] = vexpq_f32(vmulq_f32(vec_elements_flt.val[3], scale_beta_vec)); - vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vec_elements_flt.val[0]); - vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vec_elements_flt.val[1]); - vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vec_elements_flt.val[2]); - vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vec_elements_flt.val[3]); + auto vec_elements = wrapper::vloadq(in_ptr + x); + vec_elements = wrapper::vqsub(vec_max, vec_elements); + auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements); + + if (is_log) + { + vec_elements_flt.val[0] = vmulq_f32(vec_elements_flt.val[0], scale_beta_vec); + vec_elements_flt.val[1] = vmulq_f32(vec_elements_flt.val[1], scale_beta_vec); + vec_elements_flt.val[2] = vmulq_f32(vec_elements_flt.val[2], scale_beta_vec); + vec_elements_flt.val[3] = vmulq_f32(vec_elements_flt.val[3], scale_beta_vec); + vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vexpq_f32(vec_elements_flt.val[0])); + vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vexpq_f32(vec_elements_flt.val[1])); + vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vexpq_f32(vec_elements_flt.val[2])); + vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vexpq_f32(vec_elements_flt.val[3])); + } + else + { + vec_elements_flt.val[0] = vexpq_f32(vmulq_f32(vec_elements_flt.val[0], scale_beta_vec)); + vec_elements_flt.val[1] = vexpq_f32(vmulq_f32(vec_elements_flt.val[1], scale_beta_vec)); + vec_elements_flt.val[2] = vexpq_f32(vmulq_f32(vec_elements_flt.val[2], scale_beta_vec)); + vec_elements_flt.val[3] = vexpq_f32(vmulq_f32(vec_elements_flt.val[3], scale_beta_vec)); + vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vec_elements_flt.val[0]); + vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vec_elements_flt.val[1]); + vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vec_elements_flt.val[2]); + vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vec_elements_flt.val[3]); + } + + vst4q_f32(tmp_ptr + x, vec_elements_flt); } - vst4q_f32(tmp_ptr + x, vec_elements_flt); - } + /* Reduce sum */ + const auto sum_16_byte = + vaddq_f32(vaddq_f32(vec_sum.val[0], vec_sum.val[1]), vaddq_f32(vec_sum.val[2], vec_sum.val[3])); + auto sum_res = vpadd_f32(vget_high_f32(sum_16_byte), vget_low_f32(sum_16_byte)); + sum_res = vpadd_f32(sum_res, sum_res); + sum = wrapper::vgetlane(sum_res, 0); - /* Reduce sum */ - const auto sum_16_byte = vaddq_f32(vaddq_f32(vec_sum.val[0], vec_sum.val[1]), vaddq_f32(vec_sum.val[2], vec_sum.val[3])); - auto sum_res = vpadd_f32(vget_high_f32(sum_16_byte), vget_low_f32(sum_16_byte)); - sum_res = vpadd_f32(sum_res, sum_res); - sum = wrapper::vgetlane(sum_res, 0); + /* Run remaining elements */ + for (; x < input_width; ++x) + { + float element{}; + if (is_log) + { + element = (max_val - in_ptr[x]) * scale_beta; + sum += std::exp(element); + } + else + { + element = std::exp((max_val - in_ptr[x]) * scale_beta); + sum += element; + } - /* Run remaining elements */ - for(; x < input_width; ++x) - { - float element{}; - if(is_log) + tmp_ptr[x] = element; + } + + if (!is_log) { - element = (max_val - in_ptr[x]) * scale_beta; - sum += std::exp(element); + sum_inversed = 256.f / sum; } else { - element = std::exp((max_val - in_ptr[x]) * scale_beta); - sum += element; + sum = std::log(sum); } - - tmp_ptr[x] = element; } - if(!is_log) - { - sum_inversed = 256.f / sum; - } - else + /* Normalize exponentials */ { - sum = std::log(sum); - } - } - - /* Normalize exponentials */ - { - constexpr bool is_qasymm8_signed = std::is_same<T, qasymm8_signed_t>::value; - /* Loop over row and compute softmax */ - int x = 0; - for(; x <= (input_width - vec_size); x += vec_size) - { - using int_vec_type = wrapper::traits::neon_vector_t<T, 16>; - float32x4x4_t vec_in = vld4q_f32(tmp_ptr + x); - int_vec_type normalized_value{}; - if(is_log) + constexpr bool is_qasymm8_signed = std::is_same<T, qasymm8_signed_t>::value; + /* Loop over row and compute softmax */ + int x = 0; + for (; x <= (input_width - vec_size); x += vec_size) { - const float32x4x4_t sub = + using int_vec_type = wrapper::traits::neon_vector_t<T, 16>; + float32x4x4_t vec_in = vld4q_f32(tmp_ptr + x); + int_vec_type normalized_value{}; + if (is_log) { - vsubq_f32(vec_in.val[0], vdupq_n_f32(sum)), - vsubq_f32(vec_in.val[1], vdupq_n_f32(sum)), - vsubq_f32(vec_in.val[2], vdupq_n_f32(sum)), - vsubq_f32(vec_in.val[3], vdupq_n_f32(sum)), - }; - normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(sub); + const float32x4x4_t sub = { + vsubq_f32(vec_in.val[0], vdupq_n_f32(sum)), + vsubq_f32(vec_in.val[1], vdupq_n_f32(sum)), + vsubq_f32(vec_in.val[2], vdupq_n_f32(sum)), + vsubq_f32(vec_in.val[3], vdupq_n_f32(sum)), + }; + normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(sub); + } + else + { + float32x4x4_t mul = { + vmulq_f32(vec_in.val[0], vdupq_n_f32(sum_inversed)), + vmulq_f32(vec_in.val[1], vdupq_n_f32(sum_inversed)), + vmulq_f32(vec_in.val[2], vdupq_n_f32(sum_inversed)), + vmulq_f32(vec_in.val[3], vdupq_n_f32(sum_inversed)), + }; + + if (is_qasymm8_signed) + { + const auto offset_vec = wrapper::vdup_n(128.f, wrapper::traits::vector_128_tag{}); + mul.val[0] = wrapper::vsub(mul.val[0], offset_vec); + mul.val[1] = wrapper::vsub(mul.val[1], offset_vec); + mul.val[2] = wrapper::vsub(mul.val[2], offset_vec); + mul.val[3] = wrapper::vsub(mul.val[3], offset_vec); + } + + normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(mul); + } + wrapper::vstore(out_ptr + x, normalized_value); } - else + /* Run remaining elements */ + for (; x < input_width; ++x) { - float32x4x4_t mul = + if (is_log) { - vmulq_f32(vec_in.val[0], vdupq_n_f32(sum_inversed)), - vmulq_f32(vec_in.val[1], vdupq_n_f32(sum_inversed)), - vmulq_f32(vec_in.val[2], vdupq_n_f32(sum_inversed)), - vmulq_f32(vec_in.val[3], vdupq_n_f32(sum_inversed)), - }; - - if(is_qasymm8_signed) + out_ptr[x] = utils::cast::saturate_cast<T>(tmp_ptr[x] - sum); + } + else { - const auto offset_vec = wrapper::vdup_n(128.f, wrapper::traits::vector_128_tag{}); - mul.val[0] = wrapper::vsub(mul.val[0], offset_vec); - mul.val[1] = wrapper::vsub(mul.val[1], offset_vec); - mul.val[2] = wrapper::vsub(mul.val[2], offset_vec); - mul.val[3] = wrapper::vsub(mul.val[3], offset_vec); + out_ptr[x] = utils::cast::saturate_cast<T>((tmp_ptr[x] * sum_inversed) - + (is_qasymm8_signed ? 128.f : 0)); } - - normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(mul); - } - wrapper::vstore(out_ptr + x, normalized_value); - } - /* Run remaining elements */ - for(; x < input_width; ++x) - { - if(is_log) - { - out_ptr[x] = utils::cast::saturate_cast<T>(tmp_ptr[x] - sum); - } - else - { - out_ptr[x] = utils::cast::saturate_cast<T>((tmp_ptr[x] * sum_inversed) - (is_qasymm8_signed ? 128.f : 0)); } } - } - }, - in_it, max_it, out_it); + }, + in_it, max_it, out_it); } -template void neon_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 neon_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 neon_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 neon_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 |