zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-07-24 19:11:17 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix unintuitive metal kernel caching

Browse Source
This commit is contained in:
Awni Hannun 2025-06-03 08:01:53 -07:00
parent c6a20b427a
commit ac1117b224
4 changed files with 313 additions and 244 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

498
docs/src/dev/custom_metal_kernels.rst
View File

@ -8,23 +8,26 @@ MLX supports writing custom Metal kernels through the Python and C++ APIs.
Simple Example
--------------
.. currentmodule:: mlx.core
Let's write a custom kernel that computes ``exp`` elementwise:
.. code-block:: python
def exp_elementwise(a: mx.array):
source = """
uint elem = thread_position_in_grid.x;
T tmp = inp[elem];
out[elem] = metal::exp(tmp);
"""
source = """
uint elem = thread_position_in_grid.x;
T tmp = inp[elem];
out[elem] = metal::exp(tmp);
"""
kernel = mx.fast.metal_kernel(
name="myexp",
input_names=["inp"],
output_names=["out"],
source=source,
)
kernel = mx.fast.metal_kernel(
name="myexp",
input_names=["inp"],
output_names=["out"],
source=source,
)
def exp_elementwise(a: mx.array):
outputs = kernel(
inputs=[a],
template=[("T", mx.float32)],
@ -39,8 +42,13 @@ Let's write a custom kernel that computes ``exp`` elementwise:
b = exp_elementwise(a)
assert mx.allclose(b, mx.exp(a))
Every time you make a kernel, a new Metal library is created and possibly
JIT compiled. To reduce the overhead from that, build the kernel once with
:func:`fast.metal_kernel` and then use it many times.
.. note::
We are only required to pass the body of the Metal kernel in ``source``.
Only pass the body of the Metal kernel in ``source``. The function
signature is generated automatically.
The full function signature will be generated using:
@ -78,44 +86,51 @@ Putting this all together, the generated function signature for ``myexp`` is as
template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads <https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_ function.
This means we will launch ``mx.prod(grid)`` threads, subdivided into ``threadgroup`` size threadgroups.
For optimal performance, each thread group dimension should be less than or equal to the corresponding grid dimension.
Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads
<https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_
function. This means we will launch ``mx.prod(grid)`` threads, subdivided into
``threadgroup`` size threadgroups. For optimal performance, each thread group
dimension should be less than or equal to the corresponding grid dimension.
Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
Passing ``verbose=True`` to :func:`ast.metal_kernel.__call__` will print the
generated code for debugging purposes.
Using Shape/Strides
-------------------
``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
when indexing.
:func:`fast.metal_kernel` supports an argument ``ensure_row_contiguous`` which
is ``True`` by default. This will copy the array inputs if needed
before the kernel is launched to ensure that the memory layout is row
contiguous. Generally this makes writing the kernel easier, since we don't
have to worry about gaps or the ordering of the dims when indexing.
If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
input array ``a`` if any are present in ``source``.
We can then use MLX's built in indexing utils to fetch the right elements for each thread.
If we want to avoid this copy, :func:`fast.metal_kernel` automatically passes
``a_shape``, ``a_strides`` and ``a_ndim`` for each input array ``a`` if any are
present in ``source``. We can then use MLX's built in indexing utils to fetch
the right elements for each thread.
Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
Let's convert ``myexp`` above to support arbitrarily strided arrays without
relying on a copy from ``ensure_row_contiguous``:
.. code-block:: python
source = """
uint elem = thread_position_in_grid.x;
// Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
T tmp = inp[loc];
// Output arrays are always row contiguous
out[elem] = metal::exp(tmp);
"""
kernel = mx.fast.metal_kernel(
name="myexp_strided",
input_names=["inp"],
output_names=["out"],
source=source
)
def exp_elementwise(a: mx.array):
source = """
uint elem = thread_position_in_grid.x;
// Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
T tmp = inp[loc];
// Output arrays are always row contiguous
out[elem] = metal::exp(tmp);
"""
kernel = mx.fast.metal_kernel(
name="myexp_strided",
input_names=["inp"],
output_names=["out"],
source=source
)
outputs = kernel(
inputs=[a],
template=[("T", mx.float32)],
@ -142,137 +157,139 @@ We'll start with the following MLX implementation using standard ops:
.. code-block:: python
def grid_sample_ref(x, grid):
N, H_in, W_in, _ = x.shape
ix = ((grid[..., 0] + 1) * W_in - 1) / 2
iy = ((grid[..., 1] + 1) * H_in - 1) / 2
def grid_sample_ref(x, grid):
N, H_in, W_in, _ = x.shape
ix = ((grid[..., 0] + 1) * W_in - 1) / 2
iy = ((grid[..., 1] + 1) * H_in - 1) / 2
ix_nw = mx.floor(ix).astype(mx.int32)
iy_nw = mx.floor(iy).astype(mx.int32)
ix_nw = mx.floor(ix).astype(mx.int32)
iy_nw = mx.floor(iy).astype(mx.int32)
ix_ne = ix_nw + 1
iy_ne = iy_nw
ix_ne = ix_nw + 1
iy_ne = iy_nw
ix_sw = ix_nw
iy_sw = iy_nw + 1
ix_sw = ix_nw
iy_sw = iy_nw + 1
ix_se = ix_nw + 1
iy_se = iy_nw + 1
ix_se = ix_nw + 1
iy_se = iy_nw + 1
nw = (ix_se - ix) * (iy_se - iy)
ne = (ix - ix_sw) * (iy_sw - iy)
sw = (ix_ne - ix) * (iy - iy_ne)
se = (ix - ix_nw) * (iy - iy_nw)
nw = (ix_se - ix) * (iy_se - iy)
ne = (ix - ix_sw) * (iy_sw - iy)
sw = (ix_ne - ix) * (iy - iy_ne)
se = (ix - ix_nw) * (iy - iy_nw)
I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
I_nw *= mask_nw[..., None]
I_ne *= mask_ne[..., None]
I_sw *= mask_sw[..., None]
I_se *= mask_se[..., None]
I_nw *= mask_nw[..., None]
I_ne *= mask_ne[..., None]
I_sw *= mask_sw[..., None]
I_se *= mask_se[..., None]
output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
return output
return output
Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
Now let's use :func:`custom_function` together with :func:`fast.metal_kernel`
to write a fast GPU kernel for both the forward and backward passes.
First we'll implement the forward pass as a fused kernel:
.. code-block:: python
@mx.custom_function
def grid_sample(x, grid):
source = """
uint elem = thread_position_in_grid.x;
int H = x_shape[1];
int W = x_shape[2];
int C = x_shape[3];
int gH = grid_shape[1];
int gW = grid_shape[2];
assert x.ndim == 4, "`x` must be 4D."
assert grid.ndim == 4, "`grid` must be 4D."
int w_stride = C;
int h_stride = W * w_stride;
int b_stride = H * h_stride;
B, _, _, C = x.shape
_, gN, gM, D = grid.shape
out_shape = (B, gN, gM, C)
uint grid_idx = elem / C * 2;
float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
assert D == 2, "Last dim of `grid` must be size 2."
int ix_nw = floor(ix);
int iy_nw = floor(iy);
source = """
uint elem = thread_position_in_grid.x;
int H = x_shape[1];
int W = x_shape[2];
int C = x_shape[3];
int gH = grid_shape[1];
int gW = grid_shape[2];
int ix_ne = ix_nw + 1;
int iy_ne = iy_nw;
int w_stride = C;
int h_stride = W * w_stride;
int b_stride = H * h_stride;
int ix_sw = ix_nw;
int iy_sw = iy_nw + 1;
uint grid_idx = elem / C * 2;
float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
int ix_se = ix_nw + 1;
int iy_se = iy_nw + 1;
int ix_nw = floor(ix);
int iy_nw = floor(iy);
T nw = (ix_se - ix) * (iy_se - iy);
T ne = (ix - ix_sw) * (iy_sw - iy);
T sw = (ix_ne - ix) * (iy - iy_ne);
T se = (ix - ix_nw) * (iy - iy_nw);
int ix_ne = ix_nw + 1;
int iy_ne = iy_nw;
int batch_idx = elem / C / gH / gW * b_stride;
int channel_idx = elem % C;
int base_idx = batch_idx + channel_idx;
int ix_sw = ix_nw;
int iy_sw = iy_nw + 1;
T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
int ix_se = ix_nw + 1;
int iy_se = iy_nw + 1;
I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
T nw = (ix_se - ix) * (iy_se - iy);
T ne = (ix - ix_sw) * (iy_sw - iy);
T sw = (ix_ne - ix) * (iy - iy_ne);
T se = (ix - ix_nw) * (iy - iy_nw);
out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
"""
int batch_idx = elem / C / gH / gW * b_stride;
int channel_idx = elem % C;
int base_idx = batch_idx + channel_idx;
kernel = mx.fast.metal_kernel(
name="grid_sample",
input_names=["x", "grid"],
output_names=["out"],
source=source,
)
T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
@mx.custom_function
def grid_sample(x, grid):
I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
assert x.ndim == 4, "`x` must be 4D."
assert grid.ndim == 4, "`grid` must be 4D."
out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
"""
kernel = mx.fast.metal_kernel(
name="grid_sample",
input_names=["x", "grid"],
output_names=["out"],
source=source,
)
outputs = kernel(
inputs=[x, grid],
template=[("T", x.dtype)],
output_shapes=[out_shape],
output_dtypes=[x.dtype],
grid=(np.prod(out_shape), 1, 1),
threadgroup=(256, 1, 1),
)
return outputs[0]
B, _, _, C = x.shape
_, gN, gM, D = grid.shape
out_shape = (B, gN, gM, C)
assert D == 2, "Last dim of `grid` must be size 2."
outputs = kernel(
inputs=[x, grid],
template=[("T", x.dtype)],
output_shapes=[out_shape],
output_dtypes=[x.dtype],
grid=(np.prod(out_shape), 1, 1),
threadgroup=(256, 1, 1),
)
return outputs[0]
For a reasonably sized input such as:
.. code-block:: python
x.shape = (8, 1024, 1024, 64)
grid.shape = (8, 256, 256, 2)
x.shape = (8, 1024, 1024, 64)
grid.shape = (8, 256, 256, 2)
On an M1 Max, we see a big performance improvement:
@ -281,11 +298,11 @@ On an M1 Max, we see a big performance improvement:
Grid Sample VJP
---------------
Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
its custom vjp transform so MLX can differentiate it.
Since we decorated ``grid_sample`` with :func:`custom_function`, we can now
define its custom vjp transform so MLX can differentiate it.
The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
requires a few extra ``mx.fast.metal_kernel`` features:
requires a few extra :func:`fast.metal_kernel` features:
* ``init_value=0``
Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
@ -299,128 +316,129 @@ We can then implement the backwards pass as follows:
.. code-block:: python
@grid_sample.vjp
def grid_sample_vjp(primals, cotangent, _):
x, grid = primals
B, _, _, C = x.shape
_, gN, gM, D = grid.shape
source = """
uint elem = thread_position_in_grid.x;
int H = x_shape[1];
int W = x_shape[2];
int C = x_shape[3];
// Pad C to the nearest larger simdgroup size multiple
int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
assert D == 2, "Last dim of `grid` must be size 2."
int gH = grid_shape[1];
int gW = grid_shape[2];
source = """
uint elem = thread_position_in_grid.x;
int H = x_shape[1];
int W = x_shape[2];
int C = x_shape[3];
// Pad C to the nearest larger simdgroup size multiple
int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
int w_stride = C;
int h_stride = W * w_stride;
int b_stride = H * h_stride;
int gH = grid_shape[1];
int gW = grid_shape[2];
uint grid_idx = elem / C_padded * 2;
float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
int w_stride = C;
int h_stride = W * w_stride;
int b_stride = H * h_stride;
int ix_nw = floor(ix);
int iy_nw = floor(iy);
uint grid_idx = elem / C_padded * 2;
float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
int ix_ne = ix_nw + 1;
int iy_ne = iy_nw;
int ix_nw = floor(ix);
int iy_nw = floor(iy);
int ix_sw = ix_nw;
int iy_sw = iy_nw + 1;
int ix_ne = ix_nw + 1;
int iy_ne = iy_nw;
int ix_se = ix_nw + 1;
int iy_se = iy_nw + 1;
int ix_sw = ix_nw;
int iy_sw = iy_nw + 1;
T nw = (ix_se - ix) * (iy_se - iy);
T ne = (ix - ix_sw) * (iy_sw - iy);
T sw = (ix_ne - ix) * (iy - iy_ne);
T se = (ix - ix_nw) * (iy - iy_nw);
int ix_se = ix_nw + 1;
int iy_se = iy_nw + 1;
int batch_idx = elem / C_padded / gH / gW * b_stride;
int channel_idx = elem % C_padded;
int base_idx = batch_idx + channel_idx;
T nw = (ix_se - ix) * (iy_se - iy);
T ne = (ix - ix_sw) * (iy_sw - iy);
T sw = (ix_ne - ix) * (iy - iy_ne);
T se = (ix - ix_nw) * (iy - iy_nw);
T gix = T(0);
T giy = T(0);
if (channel_idx < C) {
int cot_index = elem / C_padded * C + channel_idx;
T cot = cotangent[cot_index];
if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
int batch_idx = elem / C_padded / gH / gW * b_stride;
int channel_idx = elem % C_padded;
int base_idx = batch_idx + channel_idx;
T I_nw = x[offset];
gix -= I_nw * (iy_se - iy) * cot;
giy -= I_nw * (ix_se - ix) * cot;
}
if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
T gix = T(0);
T giy = T(0);
if (channel_idx < C) {
int cot_index = elem / C_padded * C + channel_idx;
T cot = cotangent[cot_index];
if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
T I_ne = x[offset];
gix += I_ne * (iy_sw - iy) * cot;
giy -= I_ne * (ix - ix_sw) * cot;
}
if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
T I_nw = x[offset];
gix -= I_nw * (iy_se - iy) * cot;
giy -= I_nw * (ix_se - ix) * cot;
}
if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
T I_sw = x[offset];
gix -= I_sw * (iy - iy_ne) * cot;
giy += I_sw * (ix_ne - ix) * cot;
}
if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
T I_ne = x[offset];
gix += I_ne * (iy_sw - iy) * cot;
giy -= I_ne * (ix - ix_sw) * cot;
}
if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
T I_se = x[offset];
gix += I_se * (iy - iy_nw) * cot;
giy += I_se * (ix - ix_nw) * cot;
}
}
T I_sw = x[offset];
gix -= I_sw * (iy - iy_ne) * cot;
giy += I_sw * (ix_ne - ix) * cot;
}
if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
T gix_mult = W / 2;
T giy_mult = H / 2;
T I_se = x[offset];
gix += I_se * (iy - iy_nw) * cot;
giy += I_se * (ix - ix_nw) * cot;
}
}
// Reduce across each simdgroup first.
// This is much faster than relying purely on atomics.
gix = simd_sum(gix);
giy = simd_sum(giy);
T gix_mult = W / 2;
T giy_mult = H / 2;
if (thread_index_in_simdgroup == 0) {
atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
}
"""
kernel = mx.fast.metal_kernel(
name="grid_sample_grad",
input_names=["x", "grid", "cotangent"],
output_names=["x_grad", "grid_grad"],
source=source,
atomic_outputs=True,
)
// Reduce across each simdgroup first.
// This is much faster than relying purely on atomics.
gix = simd_sum(gix);
giy = simd_sum(giy);
@grid_sample.vjp
def grid_sample_vjp(primals, cotangent, _):
x, grid = primals
B, _, _, C = x.shape
_, gN, gM, D = grid.shape
if (thread_index_in_simdgroup == 0) {
atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
}
"""
kernel = mx.fast.metal_kernel(
name="grid_sample_grad",
input_names=["x", "grid", "cotangent"],
output_names=["x_grad", "grid_grad"],
source=source,
atomic_outputs=True,
)
# pad the output channels to simd group size
# so that our `simd_sum`s don't overlap.
simdgroup_size = 32
C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
grid_size = B * gN * gM * C_padded
outputs = kernel(
inputs=[x, grid, cotangent],
template=[("T", x.dtype)],
output_shapes=[x.shape, grid.shape],
output_dtypes=[x.dtype, x.dtype],
grid=(grid_size, 1, 1),
threadgroup=(256, 1, 1),
init_value=0,
)
return outputs[0], outputs[1]
assert D == 2, "Last dim of `grid` must be size 2."
# pad the output channels to simd group size
# so that our `simd_sum`s don't overlap.
simdgroup_size = 32
C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
grid_size = B * gN * gM * C_padded
outputs = kernel(
inputs=[x, grid, cotangent],
template=[("T", x.dtype)],
output_shapes=[x.shape, grid.shape],
output_dtypes=[x.dtype, x.dtype],
grid=(grid_size, 1, 1),
threadgroup=(256, 1, 1),
init_value=0,
)
return outputs[0], outputs[1]
There's an even larger speed up for the vjp:

10
mlx/backend/metal/custom_kernel.cpp
View File

@ -73,6 +73,16 @@ void CustomKernel::eval_gpu(
}
const auto [tx, ty, tz] = threadgroup_;
auto tg_size = tx * ty * tz;
auto max_tg_size = kernel->maxTotalThreadsPerThreadgroup();
if (tg_size > max_tg_size) {
std::ostringstream msg;
msg << "Thread group size (" << tg_size << ") is greater than "
<< " the maximum allowed threads per threadgroup (" << max_tg_size
<< ").";
throw std::invalid_argument(msg.str());
}
const auto [gx, gy, gz] = grid_;
MTL::Size group_dims =
MTL::Size(std::min(tx, gx), std::min(ty, gy), std::min(tz, gz));

14
mlx/fast.cpp
View File

@ -1,5 +1,6 @@
// Copyright © 2023-2024 Apple Inc.
#include <cassert>
#include <chrono>
#include <iostream>
#include <numeric>
#include <regex>
@ -1228,6 +1229,10 @@ MetalKernelFunction metal_kernel(
attributes.push_back(" " + dtype + " " + attr + " [[" + attr + "]]");
}
}
auto now = std::chrono::system_clock::now();
int64_t timestamp = std::chrono::duration_cast<std::chrono::milliseconds>(
now.time_since_epoch())
.count();
return [=,
shape_infos = std::move(shape_infos),
@ -1271,14 +1276,15 @@ MetalKernelFunction metal_kernel(
std::ostringstream func_name;
std::string template_def = "";
std::string hash_key = "";
std::string template_hash = "";
if (!template_args.empty()) {
std::regex disallowed_chars("\\<|\\>|(, )");
template_def = write_template(template_args);
hash_key = std::regex_replace(template_def, disallowed_chars, "_");
hash_key.pop_back();
template_hash = std::regex_replace(template_def, disallowed_chars, "_");
template_hash.pop_back();
}
func_name << "custom_kernel_" << name << hash_key;
func_name << "custom_kernel_" << name << "_" << template_hash << "_"
<< timestamp;
std::string kernel_name = func_name.str();
std::string kernel_source = write_signature(

35
python/tests/test_fast.py
View File

@ -735,6 +735,41 @@ class TestFast(mlx_tests.MLXTestCase):
)[0]
self.assertEqual(out.item(), 2)
@unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
def test_custom_kernel_caching(self):
def call_kernel(a: mx.array, source):
kernel = mx.fast.metal_kernel(
name="my_kernel",
input_names=["inp"],
output_names=["out"],
source=source,
)
return kernel(
inputs=[a],
grid=(a.size, 1, 1),
threadgroup=(a.size, 1, 1),
output_shapes=[a.shape],
output_dtypes=[a.dtype],
stream=mx.gpu,
)[0]
a = mx.random.normal(shape=(32,))
source = """
uint elem = thread_position_in_grid.x;
out[elem] = 0.0;
"""
out = call_kernel(a, source)
self.assertTrue(mx.array_equal(out, mx.zeros_like(out)))
source = """
uint elem = thread_position_in_grid.x;
out[elem] = 1.0;
"""
out = call_kernel(a, source)
self.assertTrue(mx.array_equal(out, mx.ones_like(out)))
if __name__ == "__main__":
unittest.main()