xmos-ai-tools 1.1.2.dev216__py3-none-macosx_11_0_arm64.whl → 1.1.2.dev236__py3-none-macosx_11_0_arm64.whl

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/sub.h

@@ -18,6 +18,7 @@ limitations under the License.
 #include <stdint.h>
 
 #include <algorithm>
+#include <cstddef>
 #include <limits>
 
 #include "ruy/profiler/instrumentation.h"  // from @ruy
@@ -29,100 +30,179 @@ namespace tflite {
 
 namespace reference_ops {
 
-inline void SubNonBroadcast(const ArithmeticParams& params,
-                            const RuntimeShape& input1_shape,
-                            const float* input1_data,
-                            const RuntimeShape& input2_shape,
-                            const float* input2_data,
-                            const RuntimeShape& output_shape,
-                            float* output_data) {
-  const int flat_size =
-      MatchingElementsSize(input1_shape, input2_shape, output_shape);
-  for (int i = 0; i < flat_size; ++i) {
-    output_data[i] = ActivationFunctionWithMinMax(
-        input1_data[i] - input2_data[i], params.float_activation_min,
-        params.float_activation_max);
+template <class T>
+struct SubImpl {
+  template <class F>
+  static void BroadcastInput1(const ArithmeticParams& params,
+                              const T* input1_data, const T* input2_data,
+                              T* output_data, size_t size, F binary_func) {
+    for (size_t c = 0; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[0], input2_data[c], params);
+    }
   }
-}
 
-inline void SubNonBroadcast(const ArithmeticParams& params,
-                            const RuntimeShape& input1_shape,
-                            const int32_t* input1_data,
-                            const RuntimeShape& input2_shape,
-                            const int32_t* input2_data,
-                            const RuntimeShape& output_shape,
-                            int32_t* output_data) {
-  const int flat_size =
-      MatchingElementsSize(input1_shape, input2_shape, output_shape);
-  for (int i = 0; i < flat_size; ++i) {
-    output_data[i] = ActivationFunctionWithMinMax(
-        input1_data[i] - input2_data[i], params.quantized_activation_min,
-        params.quantized_activation_max);
+  template <class F>
+  static void BroadcastInput2(const ArithmeticParams& params,
+                              const T* input1_data, const T* input2_data,
+                              T* output_data, size_t size, F binary_func) {
+    for (size_t c = 0; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[c], input2_data[0], params);
+    }
   }
-}
 
-// TODO(b/151345304): We can implement BroadcastSub on buffers of arbitrary
-// dimensionality if the runtime code does a single loop over one dimension
-// that handles broadcasting as the base case. The code generator would then
-// generate max(D1, D2) nested for loops.
-template <int N = 5>
-inline void BroadcastSubSlow(const ArithmeticParams& params,
-                             const RuntimeShape& input1_shape,
-                             const float* input1_data,
-                             const RuntimeShape& input2_shape,
-                             const float* input2_data,
-                             const RuntimeShape& output_shape,
-                             float* output_data) {
-  ruy::profiler::ScopeLabel label("BroadcastSubSlow/float");
-  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+  template <class F>
+  static void ElementWise(const ArithmeticParams& params, const T* input1_data,
+                          const T* input2_data, T* output_data, size_t size,
+                          F binary_func) {
+    for (size_t c = 0; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[c], input2_data[c], params);
+    }
+  }
+};
+
+template <>
+struct SubImpl<int32_t> {
+  template <class F>
+  static void BroadcastInput1(const ArithmeticParams& params,
+                              const int32_t* input1_data,
+                              const int32_t* input2_data, int32_t* output_data,
+                              size_t size, F binary_func) {
+    size_t c = 0;
+    int32_t activation_min, activation_max;
+    GetActivationParams(params, &activation_min, &activation_max);
+#ifdef USE_NEON
+    const int32x4_t vmax = vdupq_n_s32(activation_max);
+    const int32x4_t vmin = vdupq_n_s32(activation_min);
+    const int32x4_t va = vdupq_n_s32(input1_data[0]);
+    for (; c + 4 <= size; c += 4) {
+      const int32x4_t vb = vld1q_s32(&input2_data[c]);
+      int32x4_t vres = vsubq_s32(va, vb);
+      vres = vmaxq_s32(vmin, vres);
+      vres = vminq_s32(vmax, vres);
+      vst1q_s32(&output_data[c], vres);
+    }
+#endif
+    for (; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[0], input2_data[c], params);
+    }
+  }
 
-  // In Tensorflow, the dimensions are canonically named (batch_number, row,
-  // col, channel), with extents (batches, height, width, depth), with the
-  // trailing dimension changing most rapidly (channels has the smallest stride,
-  // typically 1 element).
-  //
-  // In generated C code, we store arrays with the dimensions reversed. The
-  // first dimension has smallest stride.
-  //
-  // We name our variables by their Tensorflow convention, but generate C code
-  // nesting loops such that the innermost loop has the smallest stride for the
-  // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        ActivationFunctionWithMinMax(
-            input1_data[SubscriptToIndex(desc1, indexes)] -
-                input2_data[SubscriptToIndex(desc2, indexes)],
-            params.float_activation_min, params.float_activation_max);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
+  template <class F>
+  static void BroadcastInput2(const ArithmeticParams& params,
+                              const int32_t* input1_data,
+                              const int32_t* input2_data, int32_t* output_data,
+                              size_t size, F binary_func) {
+    size_t c = 0;
+    int32_t activation_min, activation_max;
+    GetActivationParams(params, &activation_min, &activation_max);
+#ifdef USE_NEON
+    const int32x4_t vmax = vdupq_n_s32(activation_max);
+    const int32x4_t vmin = vdupq_n_s32(activation_min);
+    const int32x4_t vb = vdupq_n_s32(input2_data[0]);
+    for (; c + 4 <= size; c += 4) {
+      const int32x4_t va = vld1q_s32(&input1_data[c]);
+      int32x4_t vres = vsubq_s32(va, vb);
+      vres = vmaxq_s32(vmin, vres);
+      vres = vminq_s32(vmax, vres);
+      vst1q_s32(&output_data[c], vres);
+    }
+#endif
+    for (; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[c], input2_data[0], params);
+    }
+  }
+
+  template <class F>
+  static void ElementWise(const ArithmeticParams& params,
+                          const int32_t* input1_data,
+                          const int32_t* input2_data, int32_t* output_data,
+                          size_t size, F binary_func) {
+    size_t c = 0;
+    int32_t activation_min, activation_max;
+    GetActivationParams(params, &activation_min, &activation_max);
+#ifdef USE_NEON
+    int32x4_t vmax = vdupq_n_s32(activation_max);
+    int32x4_t vmin = vdupq_n_s32(activation_min);
+    for (; c + 4 <= size; c += 4) {
+      const int32x4_t va = vld1q_s32(&input1_data[c]);
+      const int32x4_t vb = vld1q_s32(&input2_data[c]);
+      int32x4_t vres = vsubq_s32(va, vb);
+      vres = vmaxq_s32(vmin, vres);
+      vres = vminq_s32(vmax, vres);
+      vst1q_s32(&output_data[c], vres);
+    }
+#endif
+    for (; c < size; ++c) {
+      output_data[c] = binary_func(input1_data[c], input2_data[c], params);
+    }
+  }
+};
+
+template <typename T, typename F>
+inline void BroadcastSubRecursiveDimensions(
+    int dimension, const ArithmeticParams& params, const T* input1_data,
+    const T* input2_data, T* output_data, size_t* input1_offset_p,
+    size_t* input2_offset_p, size_t* output_offset,
+    size_t* compressed_input1_stride, size_t* compressed_input2_stride,
+    size_t* compressed_output_shape, F binary_func) {
+  if (dimension > 0) {
+    for (size_t c = 0; c < compressed_output_shape[dimension]; ++c) {
+      size_t input1_offset_c = *input1_offset_p;
+      size_t input2_offset_c = *input2_offset_p;
+      BroadcastSubRecursiveDimensions(
+          dimension - 1, params, input1_data, input2_data, output_data,
+          &input1_offset_c, &input2_offset_c, output_offset,
+          compressed_input1_stride, compressed_input2_stride,
+          compressed_output_shape, binary_func);
+      *input1_offset_p += compressed_input1_stride[dimension];
+      *input2_offset_p += compressed_input2_stride[dimension];
+    }
+  } else {
+    TFLITE_DCHECK(dimension == 0);
+    bool input1_is_broadcast = compressed_input1_stride[dimension] == 0;
+    bool input2_is_broadcast = compressed_input2_stride[dimension] == 0;
+    TFLITE_DCHECK(!(input1_is_broadcast && input2_is_broadcast));
+    const T* input1_data_ptr = input1_data + *input1_offset_p;
+    const T* input2_data_ptr = input2_data + *input2_offset_p;
+    T* output_data_ptr = output_data + *output_offset;
+    if (input1_is_broadcast) {
+      // input1 is broadcast.
+      SubImpl<T>::BroadcastInput1(
+          params, input1_data_ptr, input2_data_ptr, output_data_ptr,
+          compressed_output_shape[dimension], binary_func);
+      *input2_offset_p += compressed_output_shape[dimension];
+    } else if (input2_is_broadcast) {
+      // input2 is broadcast.
+      SubImpl<T>::BroadcastInput2(
+          params, input1_data_ptr, input2_data_ptr, output_data_ptr,
+          compressed_output_shape[dimension], binary_func);
+      *input1_offset_p += compressed_output_shape[dimension];
+    } else {
+      // Add element-wise.
+      SubImpl<T>::ElementWise(params, input1_data_ptr, input2_data_ptr,
+                              output_data_ptr,
+                              compressed_output_shape[dimension], binary_func);
+      *input1_offset_p += compressed_output_shape[dimension];
+      *input2_offset_p += compressed_output_shape[dimension];
+    }
+    *output_offset += compressed_output_shape[dimension];
+  }
 }
 
-template <int N = 5>
-inline void BroadcastSubSlow(const ArithmeticParams& params,
-                             const RuntimeShape& input1_shape,
-                             const int32_t* input1_data,
-                             const RuntimeShape& input2_shape,
-                             const int32_t* input2_data,
-                             const RuntimeShape& output_shape,
-                             int32_t* output_data) {
-  ruy::profiler::ScopeLabel label("BroadcastSubSlow/int32_t");
-  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+// TODO: b/296510380 - we may be able to factor out this to common.h for all
+// binary arithmetic ops (add, sub, mul).
+template <typename T, typename F>
+inline void BroadcastSubCommon(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const T* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const T* input2_data,
+                               const RuntimeShape& output_shape, T* output_data,
+                               F binary_func) {
+  constexpr int kMaxBroadcastDim = 6;
+  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), kMaxBroadcastDim);
+  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), kMaxBroadcastDim);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), kMaxBroadcastDim);
 
   // In Tensorflow, the dimensions are canonically named (batch_number, row,
   // col, channel), with extents (batches, height, width, depth), with the
@@ -135,33 +215,6 @@ inline void BroadcastSubSlow(const ArithmeticParams& params,
   // We name our variables by their Tensorflow convention, but generate C code
   // nesting loops such that the innermost loop has the smallest stride for the
   // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        ActivationFunctionWithMinMax(
-            input1_data[SubscriptToIndex(desc1, indexes)] -
-                input2_data[SubscriptToIndex(desc2, indexes)],
-            params.quantized_activation_min, params.quantized_activation_max);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
-}
-
-template <int N = 5>
-void BroadcastSubSlow(const ArithmeticParams& params,
-                      const RuntimeShape& input1_shape,
-                      const int64_t* input1_data,
-                      const RuntimeShape& input2_shape,
-                      const int64_t* input2_data,
-                      const RuntimeShape& output_shape, int64_t* output_data) {
-  ruy::profiler::ScopeLabel label("BroadcastSubSlow/int64_t");
-  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
 
   // In Tensorflow, the dimensions are canonically named (batch_number, row,
   // col, channel), with extents (batches, height, width, depth), with the
@@ -174,54 +227,48 @@ void BroadcastSubSlow(const ArithmeticParams& params,
   // We name our variables by their Tensorflow convention, but generate C code
   // nesting loops such that the innermost loop has the smallest stride for the
   // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        ActivationFunctionWithMinMax(
-            input1_data[SubscriptToIndex(desc1, indexes)] -
-                input2_data[SubscriptToIndex(desc2, indexes)],
-            params.int64_activation_min, params.int64_activation_max);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
+
+  size_t compressed_input1_stride[kMaxBroadcastDim];
+  size_t compressed_input2_stride[kMaxBroadcastDim];
+  size_t compressed_output_shape[kMaxBroadcastDim];
+  bool broadcastable_shape = ReduceDimensionsForBroadcast<kMaxBroadcastDim>(
+      input1_shape, input2_shape, compressed_input1_stride,
+      compressed_input2_stride, compressed_output_shape);
+  // Skip broadcasting for degenerate shapes.
+  if (!broadcastable_shape) {
+    return;
+  }
+
+  size_t input1_offset = 0;
+  size_t input2_offset = 0;
+  size_t output_offset = 0;
+  BroadcastSubRecursiveDimensions(
+      kMaxBroadcastDim - 1, params, input1_data, input2_data, output_data,
+      &input1_offset, &input2_offset, &output_offset, compressed_input1_stride,
+      compressed_input2_stride, compressed_output_shape, binary_func);
 }
 
-template <typename T, int N = 5>
+// TODO(b/151345304): We can implement BroadcastSub on buffers of arbitrary
+// dimensionality if the runtime code does a single loop over one dimension
+// that handles broadcasting as the base case. The code generator would then
+// generate max(D1, D2) nested for loops.
+template <typename T>
 void BroadcastSubSlow(const ArithmeticParams& params,
                       const RuntimeShape& input1_shape, const T* input1_data,
                       const RuntimeShape& input2_shape, const T* input2_data,
                       const RuntimeShape& output_shape, T* output_data) {
-  ruy::profiler::ScopeLabel label("BroadcastSubSlow/templated");
-  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
-
-  // In Tensorflow, the dimensions are canonically named (batch_number, row,
-  // col, channel), with extents (batches, height, width, depth), with the
-  // trailing dimension changing most rapidly (channels has the smallest stride,
-  // typically 1 element).
-  //
-  // In generated C code, we store arrays with the dimensions reversed. The
-  // first dimension has smallest stride.
-  //
-  // We name our variables by their Tensorflow convention, but generate C code
-  // nesting loops such that the innermost loop has the smallest stride for the
-  // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        ActivationFunctionWithMinMax(
-            input1_data[SubscriptToIndex(desc1, indexes)] -
-                input2_data[SubscriptToIndex(desc2, indexes)],
-            params.quantized_activation_min, params.quantized_activation_max);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
+  ruy::profiler::ScopeLabel label("BroadcastSubSlow/T");
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        T activation_min, activation_max;
+        GetActivationParams(params, &activation_min, &activation_max);
+        return ActivationFunctionWithMinMax(input1_val - input2_val,
+                                            activation_min, activation_max);
+      });
 }
 
-template <int N = 5>
 inline void BroadcastSub16POTSlow(const ArithmeticParams& params,
                                   const RuntimeShape& input1_shape,
                                   const int16_t* input1_data,
@@ -230,42 +277,24 @@ inline void BroadcastSub16POTSlow(const ArithmeticParams& params,
                                   const RuntimeShape& output_shape,
                                   int16_t* output_data) {
   ruy::profiler::ScopeLabel label("BroadcastSub16POTSlow/int16_t");
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
-
-  // In Tensorflow, the dimensions are canonically named (batch_number, row,
-  // col, channel), with extents (batches, height, width, depth), with the
-  // trailing dimension changing most rapidly (channels has the smallest stride,
-  // typically 1 element).
-  //
-  // In generated C code, we store arrays with the dimensions reversed. The
-  // first dimension has smallest stride.
-  //
-  // We name our variables by their Tensorflow convention, but generate C code
-  // nesting loops such that the innermost loop has the smallest stride for the
-  // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    const int32_t input1_val = input1_data[SubscriptToIndex(desc1, indexes)];
-    const int32_t input2_val = input2_data[SubscriptToIndex(desc2, indexes)];
-    const int32_t scaled_input1_val =
-        gemmlowp::RoundingDivideByPOT(input1_val, -params.input1_shift);
-    const int32_t scaled_input2_val =
-        gemmlowp::RoundingDivideByPOT(input2_val, -params.input2_shift);
-    const int32_t raw_output = scaled_input1_val - scaled_input2_val;
-    const int32_t clamped_output =
-        std::min(params.quantized_activation_max,
-                 std::max(params.quantized_activation_min, raw_output));
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        static_cast<int16_t>(clamped_output);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
+  BroadcastSubCommon<int16_t>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](int16_t input1_val, int16_t input2_val,
+         const ArithmeticParams& params) {
+        const int32_t scaled_input1_val =
+            gemmlowp::RoundingDivideByPOT(input1_val, -params.input1_shift);
+        const int32_t scaled_input2_val =
+            gemmlowp::RoundingDivideByPOT(input2_val, -params.input2_shift);
+        const int32_t raw_output = scaled_input1_val - scaled_input2_val;
+        const int32_t clamped_output =
+            std::min(params.quantized_activation_max,
+                     std::max(params.quantized_activation_min, raw_output));
+        return static_cast<int16_t>(clamped_output);
+      });
 }
 
-template <typename T, int N = 5>
+template <typename T>
 void BroadcastQuantSubSlow(const ArithmeticParams& params,
                            const RuntimeShape& input1_shape,
                            const T* input1_data,
@@ -273,52 +302,32 @@ void BroadcastQuantSubSlow(const ArithmeticParams& params,
                            const T* input2_data,
                            const RuntimeShape& output_shape, T* output_data) {
   ruy::profiler::ScopeLabel label("BroadcastQuantSubSlow/T");
-  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
-  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
-  NdArrayDesc<N> desc1;
-  NdArrayDesc<N> desc2;
-  NdArrayDesc<N> output_desc;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
-
-  // In Tensorflow, the dimensions are canonically named (batch_number, row,
-  // col, channel), with extents (batches, height, width, depth), with the
-  // trailing dimension changing most rapidly (channels has the smallest stride,
-  // typically 1 element).
-  //
-  // In generated C code, we store arrays with the dimensions reversed. The
-  // first dimension has smallest stride.
-  //
-  // We name our variables by their Tensorflow convention, but generate C code
-  // nesting loops such that the innermost loop has the smallest stride for the
-  // best cache behavior.
-  auto sub_func = [&](int indexes[N]) {
-    const int32_t input1_val =
-        params.input1_offset + input1_data[SubscriptToIndex(desc1, indexes)];
-    const int32_t input2_val =
-        params.input2_offset + input2_data[SubscriptToIndex(desc2, indexes)];
-    const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
-    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
-    const int32_t scaled_input1_val =
-        MultiplyByQuantizedMultiplierSmallerThanOneExp(
-            shifted_input1_val, params.input1_multiplier, params.input1_shift);
-    const int32_t scaled_input2_val =
-        MultiplyByQuantizedMultiplierSmallerThanOneExp(
-            shifted_input2_val, params.input2_multiplier, params.input2_shift);
-    const int32_t raw_sub = scaled_input1_val - scaled_input2_val;
-    const int32_t raw_output =
-        MultiplyByQuantizedMultiplierSmallerThanOneExp(
-            raw_sub, params.output_multiplier, params.output_shift) +
-        params.output_offset;
-    const int32_t clamped_output =
-        std::min(params.quantized_activation_max,
-                 std::max(params.quantized_activation_min, raw_output));
-    output_data[SubscriptToIndex(output_desc, indexes)] =
-        static_cast<T>(clamped_output);
-  };
-  NDOpsHelper<N>(output_desc, sub_func);
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        const int32_t shifted_input1_val =
+            (params.input1_offset + input1_val) * (1 << params.left_shift);
+        const int32_t shifted_input2_val =
+            (params.input2_offset + input2_val) * (1 << params.left_shift);
+        const int32_t scaled_input1_val =
+            MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                shifted_input1_val, params.input1_multiplier,
+                params.input1_shift);
+        const int32_t scaled_input2_val =
+            MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                shifted_input2_val, params.input2_multiplier,
+                params.input2_shift);
+        const int32_t raw_sub = scaled_input1_val - scaled_input2_val;
+        const int32_t raw_output =
+            MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                raw_sub, params.output_multiplier, params.output_shift) +
+            params.output_offset;
+        const int32_t clamped_output =
+            std::min(params.quantized_activation_max,
+                     std::max(params.quantized_activation_min, raw_output));
+        return static_cast<T>(clamped_output);
+      });
 }
 
 // Element-wise add that can often be used for inner loop of broadcast add as
@@ -405,35 +414,12 @@ void Sub(const ArithmeticParams& params, const RuntimeShape& input1_shape,
          const T* input1_data, const RuntimeShape& input2_shape,
          const T* input2_data, const RuntimeShape& output_shape,
          T* output_data) {
-  NdArrayDesc<4> desc1;
-  NdArrayDesc<4> desc2;
-  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
-                                      &desc2);
-  const RuntimeShape extended_output_shape =
-      RuntimeShape::ExtendedShape(4, output_shape);
-
-  // In Tensorflow, the dimensions are canonically named (batch_number, row,
-  // col, channel), with extents (batches, height, width, depth), with the
-  // trailing dimension changing most rapidly (channels has the smallest stride,
-  // typically 1 element).
-  //
-  // In generated C code, we store arrays with the dimensions reversed. The
-  // first dimension has smallest stride.
-  //
-  // We name our variables by their Tensorflow convention, but generate C code
-  // nesting loops such that the innermost loop has the smallest stride for the
-  // best cache behavior.
-  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
-    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
-      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
-        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
-          output_data[Offset(extended_output_shape, b, y, x, c)] =
-              input1_data[SubscriptToIndex(desc1, b, y, x, c)] -
-              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
-        }
-      }
-    }
-  }
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        return input1_val - input2_val;
+      });
 }
 
 inline void SetActivationMinMax(const ArithmeticParams& params,

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/runtime_shape.h

@@ -15,6 +15,10 @@ limitations under the License.
 #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_RUNTIME_SHAPE_H_
 #define TENSORFLOW_LITE_KERNELS_INTERNAL_RUNTIME_SHAPE_H_
 
+#include <cstring>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+
 namespace tflite {
 
 template <int N>
@@ -34,9 +38,12 @@ class RuntimeShape {
 
   RuntimeShape() : size_(0) {}
 
-  explicit RuntimeShape(int dimensions_count) : size_(dimensions_count) {}
+  explicit RuntimeShape(int dimensions_count) : size_(dimensions_count) {
+    TFLITE_DCHECK_LE(dimensions_count, kMaxSmallSize);
+  }
 
   RuntimeShape(int shape_size, int32_t value) : size_(shape_size) {
+    TFLITE_DCHECK_LE(shape_size, kMaxSmallSize);
     for (int i = 0; i < shape_size; ++i) {
       SetDim(i, value);
     }
@@ -44,6 +51,7 @@ class RuntimeShape {
 
   RuntimeShape(int dimensions_count, const int32_t* dims_data)
       : size_(dimensions_count) {
+    // check of dimensions_count handled by ReplaceWith()
     ReplaceWith(dimensions_count, dims_data);
   }
 
@@ -69,6 +77,7 @@ class RuntimeShape {
 
   static RuntimeShape ExtendedShape(int new_shape_size,
                                     const RuntimeShape& shape) {
+    TFLITE_DCHECK_LE(new_shape_size, kMaxSmallSize);
     return RuntimeShape(new_shape_size, shape, 1);
   }
   int32_t* DimsData() { return dims_; }
@@ -76,6 +85,7 @@ class RuntimeShape {
   const int32_t* DimsDataUpTo5D() const { return dims_; }
 
   void ReplaceWith(int dimensions_count, const int32_t* dims_data) {
+    TFLITE_DCHECK_LE(dimensions_count, kMaxSmallSize);
     size_ = dimensions_count;
     int32_t* dst_dims = DimsData();
     std::memcpy(dst_dims, dims_data, dimensions_count * sizeof(int32_t));

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/strided_slice_logic.h

@@ -103,6 +103,7 @@ inline int StridedSliceEndForAxis(const tflite::StridedSliceParams& params,
   const auto shrink_axis_mask = params.shrink_axis_mask;
   const bool shrink_axis = shrink_axis_mask & (1 << axis);
   const int axis_size = input_shape.Dims(axis);
+  const bool offset = params.offset;
   if (shrink_axis) {
     if (start >= axis_size) {
       return start;
@@ -112,6 +113,9 @@ inline int StridedSliceEndForAxis(const tflite::StridedSliceParams& params,
   }
   const auto* indices = params.stop_indices;
   int end = indices[axis];
+  if (offset) {
+    end += start;
+  }
   const int32_t stride = params.strides[axis];
   const int32_t end_mask = (params.end_mask & 1 << axis);
   if (end < 0) {
@@ -246,7 +250,7 @@ inline tflite::StridedSliceParams BuildStridedSliceParams(
     int begin_mask, int end_mask, int shrink_axis_mask,
     const std::vector<int>& start_indices, const std::vector<int>& stop_indices,
     const std::vector<int>& strides) {
-  tflite::StridedSliceParams op_params;
+  tflite::StridedSliceParams op_params{};
   const int dims_count = start_indices.size();
 
   op_params.start_indices_count = dims_count;