COMPMID-1751: Remove output_3d_depth from NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
Change-Id: I1d5bc4d24059917f9ddef0873dd3043b1f2320a8
diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp
index af84d02..33a5b4a 100644
--- a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp
@@ -62,16 +62,24 @@
if(b_offset != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
- ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_row->dimension(0) != mm_result->dimension(1));
+
+ // Check if input is a 3D reinterpretation
+ const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
+
+ // Validate input
+ ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2)));
+ ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1));
TensorShape output_shape = mm_result->tensor_shape();
if(output_shape.num_dimensions() > 1)
{
+ const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
+
TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
vector_sum_row_shape.collapse_from(1);
- output_shape.collapse_from(2);
+ output_shape.collapse_from(output_batch_idx);
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[2],
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx],
"mm_result tensor must have the same number of batches of output tensor");
if(a_offset != 0)
@@ -117,6 +125,217 @@
Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
return std::make_pair(err, win);
}
+
+template <bool is_gemm3d>
+void run_offset_contribution(const Window &window,
+ ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row,
+ int32_t a_offset, int32_t b_offset, int32_t k_offset, bool slide_vector_sum_col)
+{
+ Window collapsed_window = window.collapse_if_possible(window, Window::DimZ);
+
+ const int height_input = is_gemm3d ? mm_result->info()->dimension(1) : 0;
+ const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1;
+
+ if((a_offset != 0) && (b_offset != 0) && (vector_sum_col != nullptr) && (vector_sum_row != nullptr)) // true, true
+ {
+ // Set window for vector_sum_col
+ Window win_vector_sum_col(collapsed_window);
+ win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
+ win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+ // Set window for vector_sum_row
+ Window win_vector_sum_row(collapsed_window);
+ win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
+ win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+ Iterator vector_sum_col_it(vector_sum_col, win_vector_sum_col);
+ Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row);
+ Iterator mm_result_it(mm_result, window);
+
+ const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y();
+
+ // Offset in case vector_sum_col is batched
+ const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0;
+
+ execute_window_loop(collapsed_window, [&](const Coordinates & id)
+ {
+ const int batch_id = id.z() / depth_input;
+ const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset);
+
+ // Compute the leftover term due to a_offset.
+ int32x4x4_t a_offset_term_s32 =
+ {
+ {
+ vld1q_s32(vector_sum_col_ptr + 0),
+ vld1q_s32(vector_sum_col_ptr + 4),
+ vld1q_s32(vector_sum_col_ptr + 8),
+ vld1q_s32(vector_sum_col_ptr + 12)
+ }
+ };
+
+ a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset);
+ a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset);
+ a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset);
+ a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset);
+
+ // Compute the leftover term due to b_offset.
+ int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y()
+ + (id.z() % depth_input) * height_input);
+ b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset);
+
+ // Add a_offset_term_s32 and b_offset_term_s32
+ int32x4x4_t offset_term_s32 =
+ {
+ {
+ vdupq_n_s32(k_offset),
+ vdupq_n_s32(k_offset),
+ vdupq_n_s32(k_offset),
+ vdupq_n_s32(k_offset)
+ }
+ };
+
+ offset_term_s32.val[0] = vaddq_s32(offset_term_s32.val[0], vaddq_s32(a_offset_term_s32.val[0], b_offset_term_s32));
+ offset_term_s32.val[1] = vaddq_s32(offset_term_s32.val[1], vaddq_s32(a_offset_term_s32.val[1], b_offset_term_s32));
+ offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], vaddq_s32(a_offset_term_s32.val[2], b_offset_term_s32));
+ offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], vaddq_s32(a_offset_term_s32.val[3], b_offset_term_s32));
+
+ int32x4x4_t in_s32 =
+ {
+ {
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12)
+ }
+ };
+
+ // Add the offset terms to GEMM's result
+ in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32.val[0]);
+ in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32.val[1]);
+ in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]);
+ in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]);
+
+ // Store the result with the offset contribution
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]);
+ },
+ vector_sum_col_it, vector_sum_row_it, mm_result_it);
+ }
+ else if((a_offset == 0) && (b_offset != 0) && (vector_sum_row != nullptr)) // false, true
+ {
+ ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row);
+
+ // Set window for vector_sum_row
+ Window win_vector_sum_row(collapsed_window);
+ win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
+ win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+ Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row);
+ Iterator mm_result_it(mm_result, window);
+
+ const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y();
+
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int batch_id = id.z() / depth_input;
+
+ // Compute the leftover term due to b_offset.
+ int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y()
+ + (id.z() % depth_input) * height_input);
+ b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset);
+
+ int32x4x4_t in_s32 =
+ {
+ {
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12)
+ }
+ };
+
+ // Add the offset terms to GEMM's result
+ in_s32.val[0] = vaddq_s32(in_s32.val[0], b_offset_term_s32);
+ in_s32.val[1] = vaddq_s32(in_s32.val[1], b_offset_term_s32);
+ in_s32.val[2] = vaddq_s32(in_s32.val[2], b_offset_term_s32);
+ in_s32.val[3] = vaddq_s32(in_s32.val[3], b_offset_term_s32);
+
+ // Store the result with the offset contribution
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]);
+ },
+ vector_sum_row_it, mm_result_it);
+ }
+ else if((a_offset != 0) && (b_offset == 0) && (vector_sum_col != nullptr)) // true, false
+ {
+ // Set window for vector_sum_col
+ Window win_vector_sum_col(collapsed_window);
+ win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
+ win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+ Iterator vector_sum_col_it(vector_sum_col, win_vector_sum_col);
+ Iterator mm_result_it(mm_result, window);
+
+ // Offset in case vector_sum_col is batched
+ const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0;
+
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int batch_id = id.z() / depth_input;
+ const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset);
+
+ // Compute the leftover term due to a_offset.
+ int32x4x4_t a_offset_term_s32 =
+ {
+ {
+ vld1q_s32(vector_sum_col_ptr + 0),
+ vld1q_s32(vector_sum_col_ptr + 4),
+ vld1q_s32(vector_sum_col_ptr + 8),
+ vld1q_s32(vector_sum_col_ptr + 12)
+ }
+ };
+
+ a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset);
+ a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset);
+ a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset);
+ a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset);
+
+ int32x4x4_t in_s32 =
+ {
+ {
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8),
+ vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12)
+ }
+ };
+
+ // Add the offset terms to GEMM's result
+ in_s32.val[0] = vaddq_s32(in_s32.val[0], a_offset_term_s32.val[0]);
+ in_s32.val[1] = vaddq_s32(in_s32.val[1], a_offset_term_s32.val[1]);
+ in_s32.val[2] = vaddq_s32(in_s32.val[2], a_offset_term_s32.val[2]);
+ in_s32.val[3] = vaddq_s32(in_s32.val[3], a_offset_term_s32.val[3]);
+
+ // Store the result with the offset contribution
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]);
+ vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]);
+ },
+ vector_sum_col_it, mm_result_it);
+ }
+ else // false, false
+ {
+ // No offset contribution from matrix A and matrix B
+ return;
+ }
+}
} // namespace
NEGEMMLowpOffsetContributionKernel::NEGEMMLowpOffsetContributionKernel()
@@ -177,193 +396,17 @@
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
- Window collapsed_window = window.collapse_if_possible(IKernel::window(), Window::DimZ);
+ // Check if input is a 3D reinterpretation
+ const bool reinterpret_as_3d = _vector_sum_row != nullptr
+ && _mm_result->info()->num_dimensions() > 1
+ && _mm_result->info()->tensor_shape().y() != _vector_sum_row->info()->tensor_shape().x();
- if(_a_offset != 0 && _b_offset != 0) // true, true
+ if(reinterpret_as_3d)
{
- // Set window for vector_sum_col
- Window win_vector_sum_col(collapsed_window);
- win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
- if(!_slide_vector_sum_col)
- {
- win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
- }
-
- // Set window for vector_sum_row
- Window win_vector_sum_row(collapsed_window);
- win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
- win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
- win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
- Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col);
- Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row);
- Iterator mm_result(_mm_result, window);
-
- const size_t sum_row_stride_y = _vector_sum_row->info()->strides_in_bytes().y();
-
- execute_window_loop(collapsed_window, [&](const Coordinates & id)
- {
- // Compute the leftover term due to a_offset.
- int32x4x4_t a_offset_term_s32 =
- {
- {
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 0),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 4),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 8),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 12)
- }
- };
-
- a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], _a_offset);
- a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], _a_offset);
- a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], _a_offset);
- a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], _a_offset);
-
- // Compute the leftover term due to b_offset.
- int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row.ptr() + id.z() * sum_row_stride_y) + id.y());
- b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, _b_offset);
-
- // Add a_offset_term_s32 and b_offset_term_s32
- int32x4x4_t offset_term_s32 =
- {
- {
- vdupq_n_s32(_k_offset),
- vdupq_n_s32(_k_offset),
- vdupq_n_s32(_k_offset),
- vdupq_n_s32(_k_offset)
- }
- };
-
- offset_term_s32.val[0] = vaddq_s32(offset_term_s32.val[0], vaddq_s32(a_offset_term_s32.val[0], b_offset_term_s32));
- offset_term_s32.val[1] = vaddq_s32(offset_term_s32.val[1], vaddq_s32(a_offset_term_s32.val[1], b_offset_term_s32));
- offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], vaddq_s32(a_offset_term_s32.val[2], b_offset_term_s32));
- offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], vaddq_s32(a_offset_term_s32.val[3], b_offset_term_s32));
-
- int32x4x4_t in_s32 =
- {
- {
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12)
- }
- };
-
- // Add the offset terms to GEMM's result
- in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32.val[0]);
- in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32.val[1]);
- in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]);
- in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]);
-
- // Store the result with the offset contribution
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
- },
- vector_sum_col, vector_sum_row, mm_result);
+ run_offset_contribution<true>(window, _mm_result, _vector_sum_col, _vector_sum_row, _a_offset, _b_offset, _k_offset, _slide_vector_sum_col);
}
- else if((_a_offset == 0) && (_b_offset != 0)) // false, true
+ else
{
- // Set window for vector_sum_row
- Window win_vector_sum_row(collapsed_window);
- win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
- win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
- win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
- Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row);
- Iterator mm_result(_mm_result, window);
-
- const size_t sum_row_stride_y = _vector_sum_row->info()->strides_in_bytes().y();
-
- execute_window_loop(window, [&](const Coordinates & id)
- {
- // Compute the leftover term due to b_offset.
- int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row.ptr() + id.z() * sum_row_stride_y) + id.y());
- b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, _b_offset);
-
- int32x4x4_t in_s32 =
- {
- {
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12)
- }
- };
-
- // Add the offset terms to GEMM's result
- in_s32.val[0] = vaddq_s32(in_s32.val[0], b_offset_term_s32);
- in_s32.val[1] = vaddq_s32(in_s32.val[1], b_offset_term_s32);
- in_s32.val[2] = vaddq_s32(in_s32.val[2], b_offset_term_s32);
- in_s32.val[3] = vaddq_s32(in_s32.val[3], b_offset_term_s32);
-
- // Store the result with the offset contribution
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
- },
- vector_sum_row, mm_result);
+ run_offset_contribution<false>(window, _mm_result, _vector_sum_col, _vector_sum_row, _a_offset, _b_offset, _k_offset, _slide_vector_sum_col);
}
- else if((_a_offset != 0) && (_b_offset == 0)) // true, false
- {
- // Set window for vector_sum_col
- Window win_vector_sum_col(collapsed_window);
- win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
- if(!_slide_vector_sum_col)
- {
- win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
- }
-
- Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col);
- Iterator mm_result(_mm_result, window);
-
- execute_window_loop(window, [&](const Coordinates & id)
- {
- // Compute the leftover term due to a_offset.
- int32x4x4_t a_offset_term_s32 =
- {
- {
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 0),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 4),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 8),
- vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 12)
- }
- };
-
- a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], _a_offset);
- a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], _a_offset);
- a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], _a_offset);
- a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], _a_offset);
-
- int32x4x4_t in_s32 =
- {
- {
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8),
- vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12)
- }
- };
-
- // Add the offset terms to GEMM's result
- in_s32.val[0] = vaddq_s32(in_s32.val[0], a_offset_term_s32.val[0]);
- in_s32.val[1] = vaddq_s32(in_s32.val[1], a_offset_term_s32.val[1]);
- in_s32.val[2] = vaddq_s32(in_s32.val[2], a_offset_term_s32.val[2]);
- in_s32.val[3] = vaddq_s32(in_s32.val[3], a_offset_term_s32.val[3]);
-
- // Store the result with the offset contribution
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
- vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
- },
- vector_sum_col, mm_result);
- }
- else // false, false
- {
- // No offset contribution from matrix A and matrix B
- return;
- }
-}
+}
\ No newline at end of file