Metal Kernel Writing Guide

This skill guides you through implementing Metal kernels for PyTorch operators on Apple Silicon.

Important: The goal of this skill is to use native Metal capabilities via the c10/metal/ infrastructure, NOT MPSGraph. Native Metal kernels provide better control, performance, and maintainability.

Overview

There are two workflows covered by this skill:

1. Adding new MPS support - Implementing a new operator from scratch
2. Migrating from MPSGraph - Converting existing MPSGraph-based operators to native Metal

Both workflows involve:

1. Update dispatch in aten/src/ATen/native/native_functions.yaml
2. Write Metal kernel in aten/src/ATen/native/mps/kernels/
3. Implement host-side stub in aten/src/ATen/native/mps/operations/

Step 1: Update native_functions.yaml

Location: aten/src/ATen/native/native_functions.yaml

For New Operators

Find the operator entry and add MPS dispatch:

# Simple MPS-specific implementation
- func: my_op(Tensor self) -> Tensor
  dispatch:
    CPU: my_op_cpu
    CUDA: my_op_cuda
    MPS: my_op_mps

# Shared implementation across devices (preferred for structured kernels)
- func: my_op.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
  dispatch:
    CPU, CUDA, MPS: my_op_out

# Structured kernel (preferred for new ops)
- func: my_op.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
  structured: True
  structured_inherits: TensorIteratorBase
  dispatch:
    CPU, CUDA, MPS: my_op_out

For Migrating from MPSGraph

When migrating an existing operator from MPSGraph to native Metal, consolidate the dispatch entry:

# BEFORE (MPSGraph-based, separate dispatch)
- func: atan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
  structured: True
  structured_inherits: TensorIteratorBase
  dispatch:
    CPU, CUDA: atan2_out
    MPS: atan2_out_mps  # Separate MPS implementation

# AFTER (native Metal, shared dispatch via stub)
- func: atan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
  structured: True
  structured_inherits: TensorIteratorBase
  dispatch:
    CPU, CUDA, MPS: atan2_out  # MPS now uses the same stub mechanism

Key change: Replace MPS: my_op_out_mps with adding MPS to the shared dispatch line (e.g., CPU, CUDA, MPS: my_op_out).

Update every overload. A single op typically has several native_functions.yaml entries — functional / inplace / .out, plus Tensor and Scalar variants. Each entry has its own dispatch: block, and each one must be moved over. Any entry left pointing at MPS: my_op_mps still routes that overload to the MPSGraph code, so callers can silently land on the old path depending on which overload they hit. Before declaring the migration done, grep the legacy function name and confirm no entry still references it.

Dispatch naming conventions:

• MPS: function_name_mps - MPS-specific implementation (old MPSGraph pattern)
• CPU, CUDA, MPS: function_name - Shared stub implementation (native Metal pattern)

Step 2: Implement Metal Kernel

Location: aten/src/ATen/native/mps/kernels/

Unary Kernel Pattern

// MyKernel.metal
#include <c10/metal/indexing.h>
#include <c10/metal/utils.h>
#include <metal_stdlib>

using namespace metal;
using namespace c10::metal;

// Define operation functor
struct my_op_functor {
  template <typename T>
  inline T operator()(const T x) {
    return /* your operation */;
  }
};

// Register for supported types
REGISTER_UNARY_OP(my_op, float, float);
REGISTER_UNARY_OP(my_op, half, half);
REGISTER_UNARY_OP(my_op, bfloat, bfloat);

Binary Kernel Pattern

struct my_binary_functor {
  template <typename T>
  inline T operator()(const T a, const T b) {
    return /* your operation */;
  }
};

REGISTER_BINARY_OP(my_binary, float, float);
REGISTER_BINARY_OP(my_binary, half, half);

Binary Kernel Type Registration Macros

For binary operations, use the convenience macros defined in BinaryKernel.metal:

// Floating-point types only (float, half, bfloat)
REGISTER_FLOAT_BINARY_OP(my_op);

// Integral types with float output (for math ops like atan2, copysign)
// Registers: long->float, int->float, short->float, uchar->float, char->float, bool->float
REGISTER_INT2FLOAT_BINARY_OP(my_op);

// Integral types with same-type output (for bitwise/logical ops)
// Registers: long, int, short, uchar, char, bool
REGISTER_INTEGER_BINARY_OP(my_op);

// Floating-point with opmath precision (for ops needing higher precision)
REGISTER_OPMATH_FLOAT_BINARY_OP(my_op);

Common patterns:

• Math functions (atan2, copysign, logaddexp): Use both REGISTER_FLOAT_BINARY_OP and REGISTER_INT2FLOAT_BINARY_OP
• Comparison/logical ops (maximum, minimum): Use both REGISTER_FLOAT_BINARY_OP and REGISTER_INTEGER_BINARY_OP
• Arithmetic ops (add, sub, mul): Use both REGISTER_FLOAT_BINARY_OP and REGISTER_INTEGER_BINARY_OP

Example for atan2 (supports both float and int inputs):

struct atan2_functor {
  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
  inline T operator()(const T a, const T b) {
    return static_cast<T>(precise::atan2(float(a), float(b)));
  }
  template <typename T, enable_if_t<is_integral_v<T>, bool> = true>
  inline float operator()(const T a, const T b) {
    return precise::atan2(float(a), float(b));
  }
};

REGISTER_FLOAT_BINARY_OP(atan2);
REGISTER_INT2FLOAT_BINARY_OP(atan2);

With Scalar Parameter

struct my_alpha_functor {
  template <typename T>
  inline T operator()(const T a, const T b, const T alpha) {
    return a + c10::metal::mul(alpha, b);
  }
};

REGISTER_UNARY_ALPHA_OP(my_alpha, float, float, float);
REGISTER_UNARY_ALPHA_OP(my_alpha, half, half, half);

Type-Specialized Functor

struct special_functor {
  // Floating point types
  template <typename T, enable_if_t<is_scalar_floating_point_v<T>, bool> = true>
  inline T operator()(const T x) {
    return precise::exp(x);  // Use precise math
  }

  // Integral types
  template <typename T, enable_if_t<is_scalar_integral_v<T>, bool> = true>
  inline float operator()(const T x) {
    return precise::exp(float(x));
  }

  // Complex types (float2 for cfloat, half2 for chalf)
  template <typename T, enable_if_t<is_complex_v<T>, bool> = true>
  inline T operator()(const T x) {
    // x.x = real, x.y = imaginary
    return T(/* real */, /* imag */);
  }
};

Note on complex types: Complex numbers in Metal are represented as vector types:

• c10::complex<float> maps to float2 (x = real, y = imaginary)
• c10::complex<half> maps to half2

Use is_complex_v<T> to specialize for complex types in functors.

Available c10/metal Utilities

utils.h:

• opmath_t<T> - Operation math type (half->float)
• accum_t<T> - Accumulation type for reductions
• max(), min() with NaN propagation

special_math.h:

• precise::exp(), precise::log(), precise::sqrt()
• precise::sin(), precise::cos(), precise::tan()
• erf(), erfc(), erfinv()

indexing.h:

• REGISTER_UNARY_OP(name, in_type, out_type)
• REGISTER_BINARY_OP(name, in_type, out_type)
• REGISTER_UNARY_ALPHA_OP(name, in_type, alpha_type, out_type)

Step 3: Implement Host-Side Stub

Location: aten/src/ATen/native/mps/operations/

Choose or create an appropriate file based on operation type:

• UnaryKernel.mm - Single input operations via stub dispatch
• BinaryKernel.mm - Two input operations via stub dispatch
• UnaryOps.mm / BinaryOps.mm - Legacy MPSGraph implementations (for reference)
• ReduceOps.mm - Reductions (sum, mean, max, etc.)
• Create new file for distinct operation categories

Stub Registration Pattern (Preferred for Native Metal)

For structured kernels that use the TensorIterator pattern:

// In BinaryKernel.mm (or appropriate file)

static void my_op_mps_kernel(TensorIteratorBase& iter) {
  lib.exec_binary_kernel(iter, "my_op");  // "my_op" matches the functor name in .metal
}

// Register the MPS stub - this connects to the dispatch system
REGISTER_DISPATCH(my_op_stub, &my_op_mps_kernel)

For unary operations:

static void my_unary_mps_kernel(TensorIteratorBase& iter) {
  lib.exec_unary_kernel(iter, "my_unary");
}

REGISTER_DISPATCH(my_unary_stub, &my_unary_mps_kernel)

Migration: Removing Old MPSGraph Implementation

When migrating from MPSGraph, also remove the old implementation:

1. Remove from BinaryOps.mm (or UnaryOps.mm):

   • Delete the TORCH_IMPL_FUNC(my_op_out_mps) implementation
   • Remove the corresponding #include <ATen/ops/my_op_native.h> header

2. Add to BinaryKernel.mm (or UnaryKernel.mm):

   • Add the static kernel function
   • Add the REGISTER_DISPATCH call

Step 4: Compile

After making changes, compile to verify everything builds correctly:

cd build && ninja torch_cpu

Testing

Basic operator support is already tested by test_output_match in test/test_mps.py. After implementing an operator, enable testing by removing expected failures:

1. Remove from common_mps.py

Location: torch/testing/_internal/common_mps.py

Find and remove the operator from skip/xfail lists:

# Remove entries like:
MPS_XFAILLIST = {
    "my_op": ...,  # Remove this line
}

MPS_SKIPLIST = {
    "my_op": ...,  # Remove this line
}

2. Remove from OpInfo decorators

Location: torch/testing/_internal/common_methods_invocations.py (or related files)

Remove MPS-specific decorators from the OpInfo:

OpInfo(
    "my_op",
    # Remove decorators like:
    # decorators=[skipMPS, expectedFailureMPS("reason")],
    ...
)

3. Run tests to verify

# Run the specific operator test
python test/test_mps.py -k test_output_match_my_op

# Or run full MPS test suite
python test/test_mps.py

Debugging Metal Kernels with torch.mps.compile_shader

Use torch.mps.compile_shader to JIT-compile and test individual Metal kernels in isolation. This is invaluable for debugging multi-kernel pipelines where you need to verify each stage independently.

Basic Usage

import torch

source = '''
#include <metal_stdlib>
using namespace metal;

kernel void my_kernel(
    const device float* input [[buffer(0)]],
    device float* output [[buffer(1)]],
    uint tid [[thread_position_in_grid]]) {
  output[tid] = input[tid] * 2.0;
}
'''

lib = torch.mps.compile_shader(source)

inp = torch.tensor([1.0, 2.0, 3.0], device='mps')
out = torch.zeros(3, device='mps')
lib.my_kernel(inp, out, threads=[3, 1, 1], group_size=[3, 1, 1])
torch.mps.synchronize()
print(out)  # tensor([2., 4., 6.], device='mps:0')

Dispatch Semantics

compile_shader uses dispatchThreads semantics (same as mtl_dispatch1DJob in PyTorch):

• threads=[N, 1, 1] — total number of threads (NOT threadgroups)
• group_size=[G, 1, 1] — threads per threadgroup

This differs from the dispatchThreadgroups API used by some host-side code. To match dispatchThreadgroups:MTLSizeMake(num_tgs, num_slices, 1) threadsPerThreadgroup:MTLSizeMake(TG_SIZE, 1, 1):

# Equivalent compile_shader call:
lib.kernel(args...,
    threads=[num_tgs * TG_SIZE, num_slices, 1],
    group_size=[TG_SIZE, 1, 1])

Constant Buffer Parameters

Pass scalar constants as single-element tensors:

slice_size = torch.tensor([1024], dtype=torch.int32, device='mps')
lib.my_kernel(data, output, slice_size, threads=[1024, 1, 1], group_size=[256, 1, 1])

Debugging Strategy for Multi-Kernel Pipelines

When a pipeline of kernels (e.g., histogram → prefix_sum → scatter) produces wrong results, test each kernel individually and verify its output against a Python/NumPy reference:

# 1. Run GPU kernel
lib.histogram(keys, hist, ..., threads=[N, 1, 1], group_size=[256, 1, 1])
torch.mps.synchronize()

# 2. Compute reference in Python
ref_hist = compute_histogram_cpu(keys.cpu().numpy(), ...)

# 3. Compare
assert np.array_equal(hist.cpu().numpy(), ref_hist), "Histogram mismatch!"

This isolates which kernel in the pipeline is broken, rather than debugging the entire pipeline at once.

Common Pitfalls

• Wrong threads count — threads is total threads, not threadgroups. For 5 threadgroups of 256, use threads=[1280, 1, 1].
• Threadgroup memory — compile_shader doesn't support [[threadgroup(N)]] parameters directly. If your kernel needs threadgroup memory, restructure to use threadgroup arrays declared inside the kernel body instead.

Working with TensorIterators

REGISTER_UNARY_OP / REGISTER_BINARY_OP hide the iterator plumbing. Kernels with extra params or non-elementwise layouts have to drive TensorIterator directly. The non-obvious rules:

• Pass TensorIteratorBase&, not Tensor&. Tensor& loses the offset/shape info with_32bit_indexing() produces. When the stub hands you a const TensorBase& (e.g. bernoulli_scalar_stub), build the iter inline with at::TensorIterator::borrowing_nullary_op(self) rather than const-casting.

• iter.tensor(0) returns the whole tensor for sub-iters. After with_32bit_indexing(), sub-iters still reference the original storage, so binding iter.tensor(0) naively makes every sub-iter overwrite the same prefix and leaves the tail uninitialized. Use bind_iter_tensors, which computes the per-chunk offset and binds buffer 0 to the slice. Dispatch with iter.numel(), never iter.tensor(0).numel():

  bind_iter_tensors(computeEncoder, iter, /*ntensors=*/1);
  mtl_setArgs<1>(computeEncoder, params, ..., numel);
  mtl_dispatch1DJob(computeEncoder, pso, threads);
  

• Prefer mtl_setArgs<N> over chained mtl_setBytes. mtl_setArgs<1>(encoder, a, b, c) binds a/b/c at slots 1/2/3 with the same overload resolution as the macros (std::array<long,2> → constant long2&).

• Use ceil_div. Host: at::ceil_div from <ATen/ceil_div.h>. Metal: c10::metal::ceil_div from <c10/metal/common.h> (unqualified after using namespace c10::metal;). Both wrap (a + b - 1) / b.

• IF_CONSTEXPR for Metal 3/4 portability. Metal 4 has if constexpr; Metal 3 doesn't. Use the macro from <c10/metal/common.h>, e.g. if IF_CONSTEXPR (sizeof(T) == 8) { ... }.

• Share the CPU/CUDA stub via REGISTER_MPS_DISPATCH. When the op has a DECLARE_DISPATCH stub upstream (distributions, fused ops), wire MPS in with REGISTER_MPS_DISPATCH(stub_name, &fn) instead of an MPS-specific native_functions.yaml entry. The stub already takes TensorIteratorBase&.

Working with Large Tensors

Most Metal kernels take numel as uint32_t and index in 32 bits, so anything past INT32_MAX needs host-side splitting. Drive this through TensorIterator rather than slicing manually.

Decompose via iter.with_32bit_indexing():

if (!iter.can_use_32bit_indexing()) {
  for (auto&& sub_iter : iter.with_32bit_indexing()) {
    my_kernel_impl(sub_iter, ...);
  }
  return;
}

Every yielded sub-iter satisfies can_use_32bit_indexing(), so recursion terminates after one level. The threshold is INT32_MAX (TensorIterator uses signed 32-bit offsets), not UINT32_MAX — a test just past INT32_MAX exercises the split but not the uint32_t cast; the cast itself only matters for numel > 2^32.

Use a checked cast when narrowing:

const uint32_t numel = c10::checked_convert<uint32_t>(iter.numel(), "uint32_t");

Use c10::checked_convert (<c10/util/TypeCast.h>) so wraparound becomes a TORCH_CHECK failure instead of a corrupt output.

Error Reporting from Kernels

NEVER copy results to CPU for the purposes of error checking. Any .item(), .cpu(), or other host read on a GPU tensor forces a full GPU->CPU sync that drains every in-flight op on the stream — not just the reduction you're inspecting. In a realistic pipeline this stalls the whole queue, and the cost dwarfs the actual check. Guarding the sync behind is_mps() doesn't help; the stall happens every time the op runs. Validate on-device instead and report errors through the mechanism below.

GPU code can't throw, but kernels can write into a shared error buffer that the host raises as c10::AcceleratorError on the next sync. Use TORCH_REPORT_ERROR(error_buf, ...) from <c10/metal/error.h>. Variadic arguments are concatenated; integers are formatted in base 10. Delivery is asynchronous — the failing thread keeps running, and the error surfaces whenever MPSStream::checkLastError() next runs (after a synchronize() or the next op that drains the stream), so don't rely on it for control flow inside the same dispatch. Crucially, this adds no forced sync: the error piggybacks on whatever sync the user's code already does.

Kernel side: take a device ErrorMessages* error_buf argument and call TORCH_REPORT_ERROR on the bad path. Skip the offending element afterwards so you don't also corrupt memory:

#include <c10/metal/error.h>

kernel void index_set_1d(
    device float* self,
    constant float* values,
    constant long* indices,
    constant uint& self_numel,
    device ::c10::metal::ErrorMessages* error_buf,
    uint tid [[thread_position_in_grid]]) {
  long idx = indices[tid];
  if (idx < 0 || idx >= long(self_numel)) {
    TORCH_REPORT_ERROR(
        error_buf, "index ", idx, " out of bounds for size ", long(self_numel));
    return;
  }
  self[idx] = values[tid];
}

Host side: bind the stream's error buffer as the corresponding argument. mtl_setArgs accepts it directly:

auto* stream = getCurrentMPSStream();
mtl_setArgs(encoder, self, values, indices, uint32_t(self.numel()),
            stream->getErrorBuffer());

The buffer is owned by MPSStream, sized for 30 messages, and reset after each checkLastError() drain. Only the first message is reported; later ones are kept for debugging but the AcceleratorError carries msg[0].

Checklist

• Added MPS dispatch to native_functions.yaml
• Implemented Metal kernel in kernels/
• Implemented host-side operator in operations/
• Handles empty tensors
• Handles non-contiguous tensors
• Supports required dtypes (float32, float16, bfloat16, and often complex types via float2/half2)
• Removed expected failures from torch/testing/_internal/common_mps.py
• Removed skip/xfail decorators from OpInfo (if applicable)