mlx 1.0.0

data/mlx/mlx/backend/common/compiled.cpp

@@ -0,0 +1,243 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+void print_constant(std::ostream& os, const array& x) {
+  switch (x.dtype()) {
+    case float32:
+      return print_float_constant<float>(os, x);
+    case float16:
+      return print_float_constant<float16_t>(os, x);
+    case bfloat16:
+      return print_float_constant<bfloat16_t>(os, x);
+    case float64:
+      return print_float_constant<double>(os, x);
+    case complex64:
+      return print_complex_constant<complex64_t>(os, x);
+    case int8:
+      os << static_cast<int32_t>(x.item<int8_t>());
+      return;
+    case int16:
+      return print_int_constant<int16_t>(os, x);
+    case int32:
+      return print_int_constant<int32_t>(os, x);
+    case int64:
+      return print_int_constant<int64_t>(os, x);
+    case uint8:
+      os << static_cast<uint32_t>(x.item<uint8_t>());
+      return;
+    case uint16:
+      return print_int_constant<uint16_t>(os, x);
+    case uint32:
+      return print_int_constant<uint32_t>(os, x);
+    case uint64:
+      return print_int_constant<uint64_t>(os, x);
+    case bool_:
+      os << std::boolalpha << x.item<bool>();
+      return;
+    default:
+      throw std::runtime_error("Unsupported constant type");
+  }
+}
+
+std::string get_type_string(Dtype d) {
+  switch (d) {
+    case float32:
+      return "float";
+    case float16:
+      return "float16_t";
+    case bfloat16:
+      return "bfloat16_t";
+    case float64:
+      return "double";
+    case complex64:
+      return "complex64_t";
+    case bool_:
+      return "bool";
+    case int8:
+      return "int8_t";
+    case int16:
+      return "int16_t";
+    case int32:
+      return "int32_t";
+    case int64:
+      return "int64_t";
+    case uint8:
+      return "uint8_t";
+    case uint16:
+      return "uint16_t";
+    case uint32:
+      return "uint32_t";
+    case uint64:
+      return "uint64_t";
+    default: {
+      std::ostringstream msg;
+      msg << "Unsupported compilation type " << d;
+      throw std::runtime_error(msg.str());
+    }
+  }
+}
+
+bool compiled_check_contiguity(
+    const std::vector<array>& inputs,
+    const Shape& shape) {
+  bool contiguous = true;
+  bool all_contig = true;
+  bool all_row_contig = true;
+  bool all_col_contig = true;
+  int non_scalar_inputs = 0;
+  for (const auto& x : inputs) {
+    if (is_scalar(x)) {
+      continue;
+    }
+    non_scalar_inputs++;
+    bool shape_eq = x.shape() == shape;
+    all_contig &= (x.flags().contiguous && shape_eq);
+    all_row_contig &= (x.flags().row_contiguous && shape_eq);
+    all_col_contig &= (x.flags().col_contiguous && shape_eq);
+  }
+  if (non_scalar_inputs > 1 && !all_row_contig && !all_col_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 1 && !all_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 0 && !shape.empty()) {
+    contiguous = false;
+  }
+  return contiguous;
+}
+
+void compiled_allocate_outputs(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::function<bool(size_t)>& is_constant,
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>&
+        mallocfn /* = allocator::malloc */) {
+  if (contiguous) {
+    int o = 0;
+    Strides strides;
+    size_t data_size;
+    array::Flags flags;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Correct size
+      // - Not a scalar
+      // - Donatable
+      // - Not a constant
+      if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
+          in.is_donatable() && !is_constant(i)) {
+        outputs[o++].copy_shared_buffer(in);
+      }
+      // Get representative input flags to properly set non-donated outputs
+      if (strides.empty() && in.size() == outputs[0].size()) {
+        strides = in.strides();
+        flags = in.flags();
+        data_size = in.data_size();
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(
+          mallocfn(data_size * outputs[o].itemsize()),
+          data_size,
+          strides,
+          flags);
+    }
+  } else {
+    int o = 0;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Row contiguous
+      // - Donatable
+      // - Correct size
+      // - Not a constant
+      if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
+          in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
+          !is_constant(i)) {
+        outputs[o].copy_shared_buffer(
+            in, outputs[o].strides(), in.flags(), in.data_size());
+        o++;
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(mallocfn(outputs[o].nbytes()));
+    }
+  }
+}
+
+std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
+    const std::vector<array>& inputs,
+    const array& out,
+    const std::function<bool(size_t)>& is_constant) {
+  const Shape& shape = out.shape();
+  bool contiguous = compiled_check_contiguity(inputs, shape);
+  if (contiguous) {
+    return {true, shape, {}};
+  }
+
+  std::vector<Strides> strides_vec{out.strides()};
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    // Skip constants.
+    if (is_constant(i)) {
+      continue;
+    }
+
+    // Skip scalar inputs.
+    const auto& x = inputs[i];
+    if (is_scalar(x)) {
+      continue;
+    }
+
+    // Broadcast the inputs to the output shape.
+    Strides xstrides;
+    size_t j = 0;
+    for (; j < shape.size() - x.ndim(); ++j) {
+      if (shape[j] == 1) {
+        xstrides.push_back(out.strides()[j]);
+      } else {
+        xstrides.push_back(0);
+      }
+    }
+    for (size_t i = 0; i < x.ndim(); ++i, ++j) {
+      if (x.shape(i) == 1) {
+        if (shape[j] == 1) {
+          xstrides.push_back(out.strides()[j]);
+        } else {
+          xstrides.push_back(0);
+        }
+      } else {
+        xstrides.push_back(x.strides()[i]);
+      }
+    }
+    strides_vec.push_back(std::move(xstrides));
+  }
+
+  auto tup = collapse_contiguous_dims(shape, strides_vec, INT32_MAX);
+  return {false, std::move(std::get<0>(tup)), std::move(std::get<1>(tup))};
+}
+
+bool compiled_use_large_index(
+    const std::vector<array>& inputs,
+    const std::vector<array>& outputs,
+    bool contiguous) {
+  if (contiguous) {
+    size_t max_size = 0;
+    for (const auto& in : inputs) {
+      max_size = std::max(max_size, in.data_size());
+    }
+    return max_size > UINT32_MAX;
+  } else {
+    size_t max_size = 0;
+    for (const auto& o : outputs) {
+      max_size = std::max(max_size, o.size());
+    }
+    return max_size > UINT32_MAX;
+  }
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/compiled.h

@@ -0,0 +1,77 @@
+// Copyright © 2023-2024 Apple Inc.
+#pragma once
+
+#include <functional>
+#include <iomanip>
+
+#include "mlx/array.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+inline bool is_static_cast(const Primitive& p) {
+  return (typeid(p) == typeid(Broadcast) || typeid(p) == typeid(AsType));
+}
+
+std::string get_type_string(Dtype d);
+
+template <typename T>
+void print_float_constant(std::ostream& os, const array& x) {
+  auto old_precision = os.precision();
+  if constexpr (std::is_same_v<T, double>) {
+    os << std::setprecision(std::numeric_limits<double>::digits10 + 1);
+  } else {
+    os << std::setprecision(std::numeric_limits<float>::digits10 + 1);
+  }
+  os << x.item<T>() << std::setprecision(old_precision);
+}
+
+template <typename T>
+void print_int_constant(std::ostream& os, const array& x) {
+  os << x.item<T>();
+}
+
+template <typename T>
+void print_complex_constant(std::ostream& os, const array& x) {
+  auto old_precision = os.precision();
+  T constant = x.item<T>();
+
+  os << get_type_string(x.dtype()) << "("
+     << std::setprecision(std::numeric_limits<float>::digits10 + 1)
+     << constant.real() << ", " << constant.imag() << ")"
+     << std::setprecision(old_precision);
+}
+
+void print_constant(std::ostream& os, const array& x);
+
+inline bool is_scalar(const array& x) {
+  return x.ndim() == 0;
+}
+
+// Check if we can use a contiguous operation given inputs and the output shape
+bool compiled_check_contiguity(
+    const std::vector<array>& inputs,
+    const Shape& shape);
+
+// Allocate space for the outputs possibly with input donation
+void compiled_allocate_outputs(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::function<bool(size_t)>& is_constant,
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>& mallocfn =
+        allocator::malloc);
+
+// Collapse contiguous dims ignoring scalars and constants.
+std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
+    const std::vector<array>& inputs,
+    const array& out,
+    const std::function<bool(size_t)>& is_constant);
+
+// Return whether the kernel should use large index.
+bool compiled_use_large_index(
+    const std::vector<array>& inputs,
+    const std::vector<array>& outputs,
+    bool contiguous);
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/copy.h

@@ -0,0 +1,50 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+enum class CopyType {
+  // Copy a raw scalar input into the full contiguous output
+  Scalar,
+
+  // Copy the raw input buffer contiguously into a raw output buffer of the same
+  // size
+  Vector,
+
+  // Copy the full virtual input to the full contiguous output
+  General,
+
+  // Copy the full virtual input to the full virtual output. We assume the
+  // input and output have the same shape.
+  GeneralGeneral
+};
+
+inline bool set_copy_output_data(
+    const array& in,
+    array& out,
+    CopyType ctype,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
+  if (ctype == CopyType::Vector) {
+    // If the input is donateable, we are doing a vector copy and the types
+    // have the same size, then the input buffer can hold the output.
+    if (is_donatable(in, out)) {
+      out.copy_shared_buffer(in);
+      return true;
+    } else {
+      out.set_data(
+          mallocfn(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+      return false;
+    }
+  } else {
+    out.set_data(mallocfn(out.nbytes()));
+    return false;
+  }
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/hadamard.h

@@ -0,0 +1,109 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include <map>
+
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+// From http://neilsloane.com/hadamard/
+constexpr std::string_view h12 = R"(
++-++++++++++
+--+-+-+-+-+-
++++-++----++
++---+--+-++-
++++++-++----
++-+---+--+-+
+++--+++-++--
++--++---+--+
+++----+++-++
++--+-++---+-
+++++----+++-
++-+--+-++---
+)";
+
+constexpr std::string_view h20 = R"(
++----+----++--++-++-
+-+----+---+++---+-++
+--+----+---+++-+-+-+
+---+----+---+++++-+-
+----+----++--++-++-+
+-+++++-----+--+++--+
++-+++-+---+-+--+++--
+++-++--+---+-+--+++-
++++-+---+---+-+--+++
+++++-----++--+-+--++
+--++-+-++-+-----++++
+---++-+-++-+---+-+++
++---++-+-+--+--++-++
+++---++-+----+-+++-+
+-++---++-+----+++++-
+-+--+--++-+----+----
++-+-----++-+----+---
+-+-+-+---+--+----+--
+--+-+++------+----+-
++--+--++------+----+
+)";
+
+constexpr std::string_view h28 = R"(
++------++----++-+--+-+--++--
+-+-----+++-----+-+--+-+--++-
+--+-----+++---+-+-+----+--++
+---+-----+++---+-+-+-+--+--+
+----+-----+++---+-+-+++--+--
+-----+-----++++--+-+--++--+-
+------++----++-+--+-+--++--+
+--++++-+-------++--+++-+--+-
+---++++-+-----+-++--+-+-+--+
++---+++--+----++-++--+-+-+--
+++---++---+----++-++--+-+-+-
++++---+----+----++-++--+-+-+
+++++--------+-+--++-++--+-+-
+-++++--------+++--++--+--+-+
+-+-++-++--++--+--------++++-
++-+-++--+--++--+--------++++
+-+-+-++--+--++--+----+---+++
++-+-+-++--+--+---+---++---++
+++-+-+-++--+------+--+++---+
+-++-+-+-++--+------+-++++---
++-++-+---++--+------+-++++--
+-++--++-+-++-+++----++------
++-++--++-+-++-+++-----+-----
+++-++---+-+-++-+++-----+----
+-++-++-+-+-+-+--+++-----+---
+--++-++++-+-+----+++-----+--
++--++-+-++-+-+----+++-----+-
+++--++-+-++-+-+----++------+
+)";
+
+inline const std::map<int, std::string_view> hadamard_matrices() {
+  return {{12, h12}, {20, h20}, {28, h28}};
+}
+
+inline std::pair<int, int> decompose_hadamard(int n) {
+  // n = m*2^k
+  int m = 1;
+  if (!is_power_of_2(n)) {
+    auto h_matrices = hadamard_matrices();
+    for (auto [factor, _] : h_matrices) {
+      if (n % factor == 0) {
+        m = factor;
+        n /= factor;
+        break;
+      }
+    }
+    if (m == 1) {
+      throw std::invalid_argument(
+          "[hadamard] Only supports n = m*2^k where m in (1, 12, 20, 28).");
+    }
+  }
+  if (n > (1 << 26)) {
+    throw std::invalid_argument(
+        "[hadamard] Only supports n = m*2^k where k <= 26");
+  }
+  return {n, m};
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/load.cpp

@@ -0,0 +1,57 @@
+// Copyright © 2023 Apple Inc.
+
+#include <algorithm>
+#include <utility>
+
+#include "mlx/primitives.h"
+#include "mlx/scheduler.h"
+
+namespace {
+
+template <const uint8_t scalar_size>
+void swap_endianness(uint8_t* data_bytes, size_t N) {
+  struct Elem {
+    uint8_t bytes[scalar_size];
+  };
+
+  Elem* data = reinterpret_cast<Elem*>(data_bytes);
+
+  for (size_t i = 0; i < N; i++) {
+    for (size_t j = 0; j < (scalar_size / 2); j++) {
+      std::swap(data[i].bytes[j], data[i].bytes[scalar_size - j - 1]);
+    }
+  }
+}
+
+} // namespace
+
+namespace mlx::core {
+
+void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+  auto read_task = [out_ptr = out.data<char>(),
+                    size = out.size(),
+                    itemsize = out.itemsize(),
+                    offset = offset_,
+                    reader = reader_,
+                    swap_endianness_ = swap_endianness_]() mutable {
+    reader->read(out_ptr, size * itemsize, offset);
+    if (swap_endianness_) {
+      switch (itemsize) {
+        case 2:
+          swap_endianness<2>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 4:
+          swap_endianness<4>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 8:
+          swap_endianness<8>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+      }
+    }
+  };
+  auto fut = io::thread_pool().enqueue(std::move(read_task)).share();
+  scheduler::enqueue(stream(), [fut = std::move(fut)]() { fut.wait(); });
+}
+
+} // namespace mlx::core

data/mlx/mlx/backend/common/matmul.h

@@ -0,0 +1,67 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/utils.h"
+
+#include <sstream>
+
+namespace mlx::core {
+
+inline std::tuple<Shape, Strides, Strides> collapse_batches(
+    const array& a,
+    const array& b) {
+  if (a.ndim() == 2) {
+    return {Shape{1}, Strides{0}, Strides{0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] =
+      collapse_contiguous_dims(A_bshape, std::vector{A_bstride, B_bstride});
+
+  auto a_batch_strides = batch_strides[0];
+  auto b_batch_strides = batch_strides[1];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    a_batch_strides.push_back(0);
+    b_batch_strides.push_back(0);
+  }
+
+  return std::make_tuple(batch_shape, a_batch_strides, b_batch_strides);
+}
+
+inline std::tuple<Shape, Strides, Strides, Strides>
+collapse_batches(const array& a, const array& b, const array& c) {
+  if (a.ndim() == 2) {
+    return {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+  Strides C_bstride{c.strides().begin(), c.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] = collapse_contiguous_dims(
+      A_bshape, std::vector{A_bstride, B_bstride, C_bstride});
+
+  auto A_batch_stride = batch_strides[0];
+  auto B_batch_stride = batch_strides[1];
+  auto C_batch_stride = batch_strides[2];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    A_batch_stride.push_back(0);
+    B_batch_stride.push_back(0);
+    C_batch_stride.push_back(0);
+  }
+
+  return std::make_tuple(
+      batch_shape, A_batch_stride, B_batch_stride, C_batch_stride);
+}
+
+} // namespace mlx::core