mlx 1.0.0

data/mlx/mlx/backend/common/reduce.cpp

@@ -0,0 +1,154 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/reduce.h"
+
+namespace mlx::core {
+
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes) {
+  for (int i = axes.size() - 1; i >= 0; i--) {
+    int a = axes[i];
+    shape.erase(shape.begin() + a);
+    strides.erase(strides.begin() + a);
+  }
+
+  return std::make_pair(shape, strides);
+}
+
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  auto shape = x.shape();
+  auto strides = x.strides();
+  return shapes_without_reduction_axes(
+      std::move(shape), std::move(strides), axes);
+}
+
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
+  // The data is all there and we are reducing over everything
+  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
+      x.flags().contiguous) {
+    return ContiguousAllReduce;
+  }
+
+  // Row contiguous input so the output is row contiguous
+  if (x.flags().row_contiguous) {
+    // Merge consecutive axes
+    Shape shape = {x.shape(axes[0])};
+    Strides strides = {x.strides()[axes[0]]};
+    for (int i = 1; i < axes.size(); i++) {
+      if (axes[i] - 1 == axes[i - 1] && x.shape(axes[i]) > 1) {
+        shape.back() *= x.shape(axes[i]);
+        strides.back() = x.strides()[axes[i]];
+      } else {
+        shape.push_back(x.shape(axes[i]));
+        strides.push_back(x.strides()[axes[i]]);
+      }
+    }
+
+    // Remove singleton axes from the plan
+    for (int i = shape.size() - 1; i >= 0; i--) {
+      if (shape[i] == 1) {
+        shape.erase(shape.begin() + i);
+        strides.erase(strides.begin() + i);
+      }
+    }
+
+    if (strides.back() == 1) {
+      return ReductionPlan(ContiguousReduce, shape, strides);
+    } else if (strides.back() > 1) {
+      return ReductionPlan(ContiguousStridedReduce, shape, strides);
+    }
+  }
+
+  // Let's check if we can optimize our access patterns
+  //
+  // 1. We have a reduction axis with stride 1. Simply call
+  //    GeneralContiguousReduce and be done with it.
+  // 2. We have transpositions and we are not reducing over the axis with
+  //    stride 1. However, we are reducing over an axis where everything is
+  //    contiguous in memory to the right of that axis. We can call strided
+  //    reduce and be done with it.
+  // 2. We have weird transpositions and expands. Copy the strides to the
+  //    output, then call strided reduce.
+
+  // Sort reduction axes by stride in order to merge them and figure out if we
+  // have a contiguous reduction.
+  std::vector<std::pair<int, int64_t>> reductions;
+  for (auto a : axes) {
+    if (x.shape(a) > 1) {
+      reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
+    }
+  }
+  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
+    bool a_is_zero = a.second == 0;
+    bool b_is_zero = b.second == 0;
+    return (a_is_zero != b_is_zero) ? a.second < b.second : a.second > b.second;
+  });
+  // Extract the two smallest and try to merge them in case the contiguous
+  // reduction can be bigger than just the last axis.
+  for (int i = reductions.size() - 1; i >= 1; i--) {
+    auto a = reductions[i];
+    auto b = reductions[i - 1];
+
+    // b.stride = a.shape * a.stride then a and b are contiguous
+    if (b.second == a.first * a.second) {
+      reductions.erase(reductions.begin() + i);
+      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
+    }
+  }
+
+  Shape shape;
+  Strides strides;
+  for (auto r : reductions) {
+    shape.push_back(r.first);
+    strides.push_back(r.second);
+  }
+
+  // We can call the contiguous reduction op for every weird way the input is
+  // structured in the rest of the axes.
+  if (strides.back() == 1) {
+    return ReductionPlan(GeneralContiguousReduce, shape, strides);
+  }
+
+  // Delegate to the general strided reduction op if the axes after
+  // strides.back() are contiguous.
+  if (strides.back() > 1) {
+    int64_t size = 1;
+    bool have_expand = false;
+    for (int i = x.ndim() - 1; i >= 0; i--) {
+      if (axes.back() == i) {
+        continue;
+      }
+
+      auto stride_i = x.strides()[i];
+      auto shape_i = x.shape(i);
+      if (stride_i == 0) {
+        if (shape_i == 1) {
+          continue;
+        }
+
+        have_expand = true;
+        break;
+      }
+
+      if (stride_i != size && shape_i != 1) {
+        break;
+      }
+      size *= shape_i;
+    }
+    // In the case of an expanded dimension we are being conservative and
+    // require the smallest reduction stride to be smaller than the maximum row
+    // contiguous size. The reason is that we can't easily know if the reduced
+    // axis is before or after an expanded dimension.
+    if (size > strides.back() || (size == strides.back() && !have_expand)) {
+      return ReductionPlan(GeneralStridedReduce, shape, strides);
+    }
+  }
+
+  return ReductionPlan(GeneralReduce, shape, strides);
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/reduce.h

@@ -0,0 +1,59 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+enum ReductionOpType {
+  // Self-explanatory. Read everything and produce 1 output.
+  ContiguousAllReduce,
+
+  // The input is contiguous and the last axis is reduced
+  // N1xR1xN2xR2x...xNnxRn
+  ContiguousReduce,
+
+  // The input is contiguous and the last axis is not reduced
+  // R1xN1xR2xN2x...xRnxNn
+  ContiguousStridedReduce,
+
+  // The input is not contiguous but the last axis is and it is reduced so we
+  // need to figure out the offsets but we can call the contiguous reduce after
+  // that.
+  // N3xR1xN1xR4x...xRn
+  GeneralContiguousReduce,
+
+  // The input is not contiguous but the last reduction axis and the last axis
+  // are so we need to figure out the offset but we can call the strided reduce
+  // after that.
+  GeneralStridedReduce,
+
+  // The input is not contiguous after the reduction axis and it may contain
+  // 0-stride axes or transpositions. We could copy the strides and produce a
+  // transposed outcome or we can read the input out of order and write the
+  // output in order.
+  GeneralReduce
+};
+
+struct ReductionPlan {
+  ReductionOpType type;
+  Shape shape;
+  Strides strides;
+
+  ReductionPlan(ReductionOpType type_, Shape shape_, Strides strides_)
+      : type(type_), shape(std::move(shape_)), strides(std::move(strides_)) {}
+  ReductionPlan(ReductionOpType type_) : type(type_) {}
+};
+
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
+
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes);
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes);
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/slicing.cpp

@@ -0,0 +1,71 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+std::tuple<int64_t, Strides> prepare_slice(
+    const array& in,
+    const Shape& start_indices,
+    const Shape& strides) {
+  int64_t data_offset = 0;
+  Strides inp_strides(in.ndim(), 0);
+  for (int i = 0; i < in.ndim(); ++i) {
+    data_offset += start_indices[i] * in.strides()[i];
+    inp_strides[i] = in.strides()[i] * strides[i];
+  }
+  return std::make_tuple(data_offset, inp_strides);
+}
+
+void shared_buffer_slice(
+    const array& in,
+    const Strides& out_strides,
+    int64_t data_offset,
+    size_t data_size,
+    array& out) {
+  // Compute row/col contiguity
+  auto [no_bsx_size, is_row_contiguous, is_col_contiguous] =
+      check_contiguity(out.shape(), out_strides);
+
+  auto flags = in.flags();
+  flags.row_contiguous = is_row_contiguous;
+  flags.col_contiguous = is_col_contiguous;
+  flags.contiguous = (no_bsx_size == data_size);
+
+  out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
+}
+
+void slice(
+    const array& in,
+    array& out,
+    const Shape& start_indices,
+    const Shape& strides) {
+  if (out.size() == 0) {
+    out.set_data(allocator::malloc(0));
+    return;
+  }
+
+  // Calculate out strides, initial offset
+  auto [data_offset, inp_strides] = prepare_slice(in, start_indices, strides);
+
+  // Get the location of the end based on the inp strides and out.shape()
+  int64_t low_idx = 0;
+  int64_t high_idx = 0;
+  for (int i = 0; i < inp_strides.size(); ++i) {
+    auto delta = inp_strides[i] * (out.shape()[i] - 1);
+    if (inp_strides[i] > 0) {
+      high_idx += delta;
+    } else {
+      low_idx += delta;
+    }
+  }
+  int64_t data_size = (high_idx - low_idx) + 1;
+  if (data_size < 0) {
+    std::ostringstream msg;
+    msg << "[slice] Computed invalid data size: " << data_size << ".";
+    throw std::runtime_error(msg.str());
+  }
+  shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/slicing.h

@@ -0,0 +1,20 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+std::tuple<int64_t, Strides> prepare_slice(
+    const array& in,
+    const Shape& start_indices,
+    const Shape& strides);
+
+void slice(
+    const array& in,
+    array& out,
+    const Shape& start_indices,
+    const Shape& strides);
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/ternary.h

@@ -0,0 +1,85 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+#include "mlx/allocator.h"
+#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+// TODO: Add support for more combinations of input types.
+enum class TernaryOpType {
+  ScalarScalarScalar,
+  VectorVectorVector,
+  VectorVectorScalar,
+  VectorScalarVector,
+  General,
+};
+
+inline TernaryOpType
+get_ternary_op_type(const array& a, const array& b, const array& c) {
+  TernaryOpType topt;
+  if (a.data_size() == 1 && b.data_size() == 1 && c.data_size() == 1) {
+    topt = TernaryOpType::ScalarScalarScalar;
+  } else if (
+      (a.flags().row_contiguous && b.flags().row_contiguous &&
+       c.flags().row_contiguous) ||
+      (a.flags().col_contiguous && b.flags().col_contiguous &&
+       c.flags().col_contiguous)) {
+    topt = TernaryOpType::VectorVectorVector;
+  } else if (
+      b.data_size() == 1 && a.flags().row_contiguous &&
+      c.flags().row_contiguous) {
+    topt = TernaryOpType::VectorScalarVector;
+  } else if (
+      c.data_size() == 1 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    topt = TernaryOpType::VectorVectorScalar;
+  } else {
+    topt = TernaryOpType::General;
+  }
+  return topt;
+}
+
+inline void set_ternary_op_output_data(
+    const array& a,
+    const array& b,
+    const array& c,
+    array& out,
+    TernaryOpType topt,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
+  auto maybe_donate = [&out](const array& x) {
+    if (is_donatable(x, out)) {
+      out.copy_shared_buffer(x);
+      return true;
+    }
+    return false;
+  };
+
+  switch (topt) {
+    case TernaryOpType::ScalarScalarScalar:
+      out.set_data(mallocfn(out.itemsize()), 1, b.strides(), b.flags());
+      break;
+    case TernaryOpType::VectorVectorVector:
+      if (!(maybe_donate(a) || maybe_donate(b) || maybe_donate(c))) {
+        out.set_data(
+            mallocfn(out.itemsize() * b.data_size()),
+            b.data_size(),
+            b.strides(),
+            b.flags());
+      }
+      break;
+    case TernaryOpType::VectorVectorScalar:
+    case TernaryOpType::VectorScalarVector:
+    case TernaryOpType::General:
+      // Try to donate an input which is row_contiguous
+      if (!((a.flags().row_contiguous && maybe_donate(a)) ||
+            (b.flags().row_contiguous && maybe_donate(b)) ||
+            (c.flags().row_contiguous && maybe_donate(c)))) {
+        out.set_data(mallocfn(out.nbytes()));
+      }
+      break;
+  }
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/unary.h

@@ -0,0 +1,29 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+inline void set_unary_output_data(
+    const array& in,
+    array& out,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
+  if (in.flags().contiguous) {
+    if (is_donatable(in, out)) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(
+          mallocfn(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    out.set_data(mallocfn(out.nbytes()));
+  }
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/utils.cpp

@@ -0,0 +1,231 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <dlfcn.h>
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
+}
+
+std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
+    const Shape& shape,
+    const std::vector<Strides>& strides,
+    int64_t size_cap) {
+  // Make a vector that has axes separated with -1. Collapse all axes between
+  // -1.
+  Shape to_collapse;
+  if (shape.size() > 0) {
+    if (shape[0] != 1) {
+      to_collapse.push_back(0);
+    }
+    size_t size = shape[0];
+    for (int i = 1; i < shape.size(); i++) {
+      bool contiguous = true;
+      size *= shape[i];
+      for (const auto& st : strides) {
+        if (st[i] * shape[i] != st[i - 1] || size > size_cap) {
+          contiguous = false;
+          size = shape[i];
+          break;
+        }
+      }
+      if (!contiguous) {
+        to_collapse.push_back(-1);
+      }
+      if (shape[i] != 1) {
+        to_collapse.push_back(i);
+      }
+    }
+    to_collapse.push_back(-1);
+  }
+
+  Shape out_shape;
+  std::vector<Strides> out_strides(strides.size());
+  for (int i = 0;;) {
+    while (i < to_collapse.size() && to_collapse[i] == -1) {
+      ++i;
+    };
+    if (i == to_collapse.size()) {
+      break;
+    }
+    int current_shape = shape[to_collapse[i]];
+    int k = i;
+    while (to_collapse[++k] != -1) {
+      current_shape *= shape[to_collapse[k]];
+    }
+    out_shape.push_back(current_shape);
+    for (int j = 0; j < strides.size(); j++) {
+      const auto& st = strides[j];
+      out_strides[j].push_back(st[to_collapse[k - 1]]);
+    }
+    i = k + 1;
+  }
+
+  if (!shape.empty() && out_shape.empty()) {
+    out_shape.push_back(1);
+    for (auto& out_stride : out_strides) {
+      out_stride.push_back(0);
+    }
+  }
+  return std::make_tuple(out_shape, out_strides);
+}
+
+std::pair<Shape, Strides> collapse_contiguous_dims(
+    const Shape& shape,
+    const Strides& strides,
+    int64_t size_cap) {
+  Shape collapsed_shape;
+  Strides collapsed_strides;
+
+  if (shape.size() > 0) {
+    collapsed_shape.push_back(shape[0]);
+    collapsed_strides.push_back(strides[0]);
+    for (int i = 1; i < shape.size(); i++) {
+      if (shape[i] == 1) {
+        continue;
+      } else if (
+          strides[i] * shape[i] != collapsed_strides.back() ||
+          collapsed_shape.back() * static_cast<int64_t>(shape[i]) > size_cap) {
+        collapsed_shape.push_back(shape[i]);
+        collapsed_strides.push_back(strides[i]);
+      } else {
+        collapsed_shape.back() *= shape[i];
+        collapsed_strides.back() = strides[i];
+      }
+    }
+  }
+
+  return std::make_pair(collapsed_shape, collapsed_strides);
+}
+
+std::pair<Shape, Strides> collapse_contiguous_dims(
+    const array& a,
+    int64_t size_cap /* = std::numeric_limits<int32_t>::max()*/) {
+  return collapse_contiguous_dims(a.shape(), a.strides(), size_cap);
+}
+
+Dims get_block_dims_common(int dim0, int dim1, int dim2, int pow2 /* = 10 */) {
+  int pows[3] = {0, 0, 0};
+  int sum = 0;
+  while (true) {
+    int presum = sum;
+    // Check all the pows
+    if (dim0 >= (1 << (pows[0] + 1))) {
+      pows[0]++;
+      sum++;
+    }
+    if (sum == 10) {
+      break;
+    }
+    if (dim1 >= (1 << (pows[1] + 1))) {
+      pows[1]++;
+      sum++;
+    }
+    if (sum == 10) {
+      break;
+    }
+    if (dim2 >= (1 << (pows[2] + 1))) {
+      pows[2]++;
+      sum++;
+    }
+    if (sum == presum || sum == pow2) {
+      break;
+    }
+  }
+  return std::make_tuple(1ul << pows[0], 1ul << pows[1], 1ul << pows[2]);
+}
+
+Dims get_2d_grid_dims_common(const Shape& shape, const Strides& strides) {
+  // Dims with strides of 0 are ignored as they
+  // correspond to broadcasted dimensions
+  size_t grid_x = 1;
+  size_t grid_y = 1;
+  for (int i = 0; i < shape.size(); ++i) {
+    if (strides[i] == 0) {
+      continue;
+    }
+    if (grid_x * shape[i] < UINT32_MAX) {
+      grid_x *= shape[i];
+    } else {
+      grid_y *= shape[i];
+    }
+  }
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
+    throw std::runtime_error("Unable to safely factor shape.");
+  }
+  if (grid_y > grid_x) {
+    std::swap(grid_x, grid_y);
+  }
+  return std::make_tuple(
+      static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
+}
+
+Dims get_2d_grid_dims_common(
+    const Shape& shape,
+    const Strides& strides,
+    size_t divisor) {
+  // Compute the 2d grid dimensions such that the total size of the grid is
+  // divided by divisor.
+  size_t grid_x = 1;
+  size_t grid_y = 1;
+  for (int i = 0; i < shape.size(); ++i) {
+    if (strides[i] == 0) {
+      continue;
+    }
+
+    // No need to add this shape we can just remove it from the divisor.
+    if (divisor % shape[i] == 0) {
+      divisor /= shape[i];
+      continue;
+    }
+
+    if (grid_x * shape[i] < UINT32_MAX) {
+      grid_x *= shape[i];
+    } else {
+      grid_y *= shape[i];
+    }
+
+    if (divisor > 1) {
+      if (grid_x % divisor == 0) {
+        grid_x /= divisor;
+        divisor = 1;
+      } else if (grid_y % divisor == 0) {
+        grid_y /= divisor;
+        divisor = 1;
+      }
+    }
+  }
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
+    throw std::runtime_error("Unable to safely factor shape.");
+  }
+  if (grid_y > grid_x) {
+    std::swap(grid_x, grid_y);
+  }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
+  return std::make_tuple(
+      static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
+}
+
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
+  auto [bx, by, bz] = get_block_dims_common(dim0, dim1, dim2);
+  auto gx = (dim0 + bx - 1) / bx;
+  auto gy = (dim1 + by - 1) / by;
+  auto gz = (dim2 + bz - 1) / bz;
+
+  return std::make_pair(
+      std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
+}
+
+} // namespace mlx::core