diff options
author | Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-27 17:46:17 +0100 |
---|---|---|
committer | felixjohnny.thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-28 12:08:05 +0000 |
commit | afd38f0c617d6f89b2b4532c6c44f116617e2b6f (patch) | |
tree | 03bc7d5a762099989b16a656fa8d397b490ed70e /src/cpu/kernels/conv3d/neon/quantized.h | |
parent | bdcb4c148ee2fdeaaddf4cf1e57bbb0de02bb894 (diff) | |
download | ComputeLibrary-afd38f0c617d6f89b2b4532c6c44f116617e2b6f.tar.gz |
Apply clang-format on repository
Code is formatted as per a revised clang format configuration
file(not part of this delivery). Version 14.0.6 is used.
Exclusion List:
- files with .cl extension
- files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...)
And the following directories
- compute_kernel_writer/validation/
- tests/
- include/
- src/core/NEON/kernels/convolution/
- src/core/NEON/kernels/arm_gemm/
- src/core/NEON/kernels/arm_conv/
- data/
There will be a follow up for formatting of .cl files and the
files under tests/ and compute_kernel_writer/validation/.
Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Diffstat (limited to 'src/cpu/kernels/conv3d/neon/quantized.h')
-rw-r--r-- | src/cpu/kernels/conv3d/neon/quantized.h | 257 |
1 files changed, 138 insertions, 119 deletions
diff --git a/src/cpu/kernels/conv3d/neon/quantized.h b/src/cpu/kernels/conv3d/neon/quantized.h index a8165b4944..f0fc9b5a71 100644 --- a/src/cpu/kernels/conv3d/neon/quantized.h +++ b/src/cpu/kernels/conv3d/neon/quantized.h @@ -28,16 +28,22 @@ #include "arm_compute/core/utils/misc/Traits.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/FunctionDescriptors.h" + +#include "src/core/helpers/WindowHelpers.h" #include "src/core/NEON/NEAsymm.h" #include "src/core/NEON/wrapper/wrapper.h" -#include "src/core/helpers/WindowHelpers.h" namespace arm_compute { namespace cpu { template <typename T> -void directconv3d_quantized_neon_ndhwc(const ITensor *src0, const ITensor *src1, const ITensor *src2, ITensor *dst, const Conv3dInfo &conv_info, const Window &window) +void directconv3d_quantized_neon_ndhwc(const ITensor *src0, + const ITensor *src1, + const ITensor *src2, + ITensor *dst, + const Conv3dInfo &conv_info, + const Window &window) { const ITensor *src = src0; const ITensor *weights = src1; @@ -104,153 +110,166 @@ void directconv3d_quantized_neon_ndhwc(const ITensor *src0, const ITensor *src1, Iterator wei(weights, window_w); const int32_t *biases_ptr = nullptr; - if(biases != nullptr) + if (biases != nullptr) { biases_ptr = reinterpret_cast<int32_t *>(biases->buffer() + biases->info()->offset_first_element_in_bytes()); } - execute_window_loop(window_out, [&](const Coordinates & id) - { - // We are computing the theoretical input starting points - const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; - const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; - const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front; - const int in_w_end_t = in_w_start_t + kernel_dim_w; - const int in_h_end_t = in_h_start_t + kernel_dim_h; - const int in_d_end_t = in_d_start_t + kernel_dim_d; + execute_window_loop( + window_out, + [&](const Coordinates &id) + { + // We are computing the theoretical input starting points + const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; + const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; + const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front; + const int in_w_end_t = in_w_start_t + kernel_dim_w; + const int in_h_end_t = in_h_start_t + kernel_dim_h; + const int in_d_end_t = in_d_start_t + kernel_dim_d; - // We are computing the valid initial and ending input points by checking the borders - const int in_w_start = std::max(in_w_start_t, 0); - const int in_h_start = std::max(in_h_start_t, 0); - const int in_d_start = std::max(in_d_start_t, 0); - const int in_w_end = std::min(in_w_end_t, input_dim_w); - const int in_h_end = std::min(in_h_end_t, input_dim_h); - const int in_d_end = std::min(in_d_end_t, input_dim_d); + // We are computing the valid initial and ending input points by checking the borders + const int in_w_start = std::max(in_w_start_t, 0); + const int in_h_start = std::max(in_h_start_t, 0); + const int in_d_start = std::max(in_d_start_t, 0); + const int in_w_end = std::min(in_w_end_t, input_dim_w); + const int in_h_end = std::min(in_h_end_t, input_dim_h); + const int in_d_end = std::min(in_d_end_t, input_dim_d); - // We use the input points to select the valid weight points to use - const int wei_w_start = in_w_start - in_w_start_t; - const int wei_h_start = in_h_start - in_h_start_t; - const int wei_d_start = in_d_start - in_d_start_t; - const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end); - const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); - const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end); + // We use the input points to select the valid weight points to use + const int wei_w_start = in_w_start - in_w_start_t; + const int wei_h_start = in_h_start - in_h_start_t; + const int wei_d_start = in_d_start - in_d_start_t; + const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end); + const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); + const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end); - const int index_c_out_end = weights->info()->dimension(0); - const int index_c_in_end = weights->info()->dimension(1); - const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[4] * input_stride_n; + const int index_c_out_end = weights->info()->dimension(0); + const int index_c_in_end = weights->info()->dimension(1); + const T *const in_ptr_start = + reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + + id[4] * input_stride_n; - execute_window_loop(window_w, [&](const Coordinates & id_w) - { - /* + execute_window_loop( + window_w, + [&](const Coordinates &id_w) + { + /* * This is the loop in the weights, and it goes along OFM (output feature map) */ - const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); - int32_t acc = static_cast<int32_t>(0); - T *out_ptr = reinterpret_cast<T *>(out.ptr()); - for(int index_wei_d = wei_d_start, index_in_d = in_d_start; index_wei_d < wei_d_end; ++index_wei_d, ++index_in_d) - { - const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d; - const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d; - for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h) - { - const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h; - const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h; - for(int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; ++index_wei_w, ++index_in_w) + const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); + int32_t acc = static_cast<int32_t>(0); + T *out_ptr = reinterpret_cast<T *>(out.ptr()); + for (int index_wei_d = wei_d_start, index_in_d = in_d_start; index_wei_d < wei_d_end; + ++index_wei_d, ++index_in_d) { - const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w; - const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w; - int index_c_in = 0; - vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); - - q32x4_t acc_q32_0 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); - q32x4_t acc_q32_1 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); - q32x4_t acc_q32_2 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); - q32x4_t acc_q32_3 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); - - for(; index_c_in <= index_c_in_end - num_elems_read_per_iteration; - index_c_in += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration) + const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d; + const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d; + for (int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; + ++index_wei_h, ++index_in_h) { - const auto src_vec = wrapper::vloadq(in_ptr_mover); - //Load Cin weights - for(int k = 0; k < num_elems_read_per_iteration; ++k, weights_ptr_mover += index_c_out_end) + const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h; + const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h; + for (int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; + ++index_wei_w, ++index_in_w) { - w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k); - } - q32x4_t src_q32_0 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); - q32x4_t src_q32_1 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); - q32x4_t src_q32_2 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); - q32x4_t src_q32_3 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); + const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w; + const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w; + int index_c_in = 0; + vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); - q32x4_t wei_q32_0 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); - q32x4_t wei_q32_1 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); - q32x4_t wei_q32_2 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); - q32x4_t wei_q32_3 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); + q32x4_t acc_q32_0 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); + q32x4_t acc_q32_1 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); + q32x4_t acc_q32_2 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); + q32x4_t acc_q32_3 = wrapper::vdup_n(static_cast<q32_t>(0), tag_type()); - const auto src_q16_0 = wrapper::vmovl(wrapper::vgetlow(src_vec)); - const auto src_q16_1 = wrapper::vmovl(wrapper::vgethigh(src_vec)); - const auto wei_q16_0 = wrapper::vmovl(wrapper::vgetlow(w_vec)); - const auto wei_q16_1 = wrapper::vmovl(wrapper::vgethigh(w_vec)); + for (; index_c_in <= index_c_in_end - num_elems_read_per_iteration; + index_c_in += num_elems_read_per_iteration, + in_ptr_mover += num_elems_read_per_iteration) + { + const auto src_vec = wrapper::vloadq(in_ptr_mover); + //Load Cin weights + for (int k = 0; k < num_elems_read_per_iteration; + ++k, weights_ptr_mover += index_c_out_end) + { + w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k); + } + q32x4_t src_q32_0 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); + q32x4_t src_q32_1 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); + q32x4_t src_q32_2 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); + q32x4_t src_q32_3 = wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type()); - src_q32_0 = wrapper::vadd(src_q32_0, wrapper::vmovl(wrapper::vgetlow(src_q16_0))); - src_q32_1 = wrapper::vadd(src_q32_1, wrapper::vmovl(wrapper::vgethigh(src_q16_0))); - src_q32_2 = wrapper::vadd(src_q32_2, wrapper::vmovl(wrapper::vgetlow(src_q16_1))); - src_q32_3 = wrapper::vadd(src_q32_3, wrapper::vmovl(wrapper::vgethigh(src_q16_1))); + q32x4_t wei_q32_0 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); + q32x4_t wei_q32_1 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); + q32x4_t wei_q32_2 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); + q32x4_t wei_q32_3 = wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type()); - wei_q32_0 = wrapper::vadd(wei_q32_0, wrapper::vmovl(wrapper::vgetlow(wei_q16_0))); - wei_q32_1 = wrapper::vadd(wei_q32_1, wrapper::vmovl(wrapper::vgethigh(wei_q16_0))); - wei_q32_2 = wrapper::vadd(wei_q32_2, wrapper::vmovl(wrapper::vgetlow(wei_q16_1))); - wei_q32_3 = wrapper::vadd(wei_q32_3, wrapper::vmovl(wrapper::vgethigh(wei_q16_1))); + const auto src_q16_0 = wrapper::vmovl(wrapper::vgetlow(src_vec)); + const auto src_q16_1 = wrapper::vmovl(wrapper::vgethigh(src_vec)); + const auto wei_q16_0 = wrapper::vmovl(wrapper::vgetlow(w_vec)); + const auto wei_q16_1 = wrapper::vmovl(wrapper::vgethigh(w_vec)); - acc_q32_0 = wrapper::vmla(acc_q32_0, wei_q32_0, src_q32_0); - acc_q32_1 = wrapper::vmla(acc_q32_1, wei_q32_1, src_q32_1); - acc_q32_2 = wrapper::vmla(acc_q32_2, wei_q32_2, src_q32_2); - acc_q32_3 = wrapper::vmla(acc_q32_3, wei_q32_3, src_q32_3); - } + src_q32_0 = wrapper::vadd(src_q32_0, wrapper::vmovl(wrapper::vgetlow(src_q16_0))); + src_q32_1 = wrapper::vadd(src_q32_1, wrapper::vmovl(wrapper::vgethigh(src_q16_0))); + src_q32_2 = wrapper::vadd(src_q32_2, wrapper::vmovl(wrapper::vgetlow(src_q16_1))); + src_q32_3 = wrapper::vadd(src_q32_3, wrapper::vmovl(wrapper::vgethigh(src_q16_1))); + + wei_q32_0 = wrapper::vadd(wei_q32_0, wrapper::vmovl(wrapper::vgetlow(wei_q16_0))); + wei_q32_1 = wrapper::vadd(wei_q32_1, wrapper::vmovl(wrapper::vgethigh(wei_q16_0))); + wei_q32_2 = wrapper::vadd(wei_q32_2, wrapper::vmovl(wrapper::vgetlow(wei_q16_1))); + wei_q32_3 = wrapper::vadd(wei_q32_3, wrapper::vmovl(wrapper::vgethigh(wei_q16_1))); + + acc_q32_0 = wrapper::vmla(acc_q32_0, wei_q32_0, src_q32_0); + acc_q32_1 = wrapper::vmla(acc_q32_1, wei_q32_1, src_q32_1); + acc_q32_2 = wrapper::vmla(acc_q32_2, wei_q32_2, src_q32_2); + acc_q32_3 = wrapper::vmla(acc_q32_3, wei_q32_3, src_q32_3); + } #if defined(__aarch64__) - acc += wrapper::vaddv(acc_q32_0); - acc += wrapper::vaddv(acc_q32_1); - acc += wrapper::vaddv(acc_q32_2); - acc += wrapper::vaddv(acc_q32_3); + acc += wrapper::vaddv(acc_q32_0); + acc += wrapper::vaddv(acc_q32_1); + acc += wrapper::vaddv(acc_q32_2); + acc += wrapper::vaddv(acc_q32_3); #else // __aarch64__ - auto temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_0), wrapper::vgetlow(acc_q32_0)); - temp = wrapper::vpadd(temp, temp); - acc += wrapper::vgetlane(temp, 0); + auto temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_0), wrapper::vgetlow(acc_q32_0)); + temp = wrapper::vpadd(temp, temp); + acc += wrapper::vgetlane(temp, 0); - temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_1), wrapper::vgetlow(acc_q32_1)); - temp = wrapper::vpadd(temp, temp); - acc += wrapper::vgetlane(temp, 0); + temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_1), wrapper::vgetlow(acc_q32_1)); + temp = wrapper::vpadd(temp, temp); + acc += wrapper::vgetlane(temp, 0); - temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_2), wrapper::vgetlow(acc_q32_2)); - temp = wrapper::vpadd(temp, temp); - acc += wrapper::vgetlane(temp, 0); + temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_2), wrapper::vgetlow(acc_q32_2)); + temp = wrapper::vpadd(temp, temp); + acc += wrapper::vgetlane(temp, 0); - temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_3), wrapper::vgetlow(acc_q32_3)); - temp = wrapper::vpadd(temp, temp); - acc += wrapper::vgetlane(temp, 0); + temp = wrapper::vpadd(wrapper::vgethigh(acc_q32_3), wrapper::vgetlow(acc_q32_3)); + temp = wrapper::vpadd(temp, temp); + acc += wrapper::vgetlane(temp, 0); #endif // __aarch64__ - for(; index_c_in < index_c_in_end; ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end) - { - const auto src_val = *(in_ptr_mover) + input_offset; - const auto w_val = *(weights_ptr_mover) + weights_offset; - acc += src_val * w_val; + for (; index_c_in < index_c_in_end; + ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end) + { + const auto src_val = *(in_ptr_mover) + input_offset; + const auto w_val = *(weights_ptr_mover) + weights_offset; + acc += src_val * w_val; + } + } } } - } - } - if(biases) - { - acc += *reinterpret_cast<const int32_t *>(biases_ptr + id_w[0]); - } + if (biases) + { + acc += *reinterpret_cast<const int32_t *>(biases_ptr + id_w[0]); + } - T out_val = finalize_quantization(acc, output_multiplier, output_shift, output_offset, T(0), T(0), false); - *(reinterpret_cast<T *>(out_ptr + id_w[0])) = out_val; + T out_val = + finalize_quantization(acc, output_multiplier, output_shift, output_offset, T(0), T(0), false); + *(reinterpret_cast<T *>(out_ptr + id_w[0])) = out_val; + }, + wei); }, - wei); - }, - out); + out); } } // namespace cpu } // namespace arm_compute -#endif // SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H
\ No newline at end of file +#endif // SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H |