Re-organize SLL ops, pt 7

fbgemm_gpu/fbgemm_gpu/sll/__init__.py

@@ -34,12 +34,7 @@
 )
 
 from fbgemm_gpu.sll.triton_sll import (  # noqa F401
-    array_jagged_bmm_jagged_out,
-    dense_jagged_cat_jagged_out,
-    jagged2_to_padded_dense,
-    # jagged_dense_bmm,
     jagged_dense_elementwise_mul_jagged_out,
-    jagged_jagged_bmm_jagged_out,
     triton_jagged_self_substraction_jagged_out,
 )
 
@@ -269,28 +264,13 @@
 
 # pyre-ignore[5]
 sll_gpu_registrations = {
-    "sll_dense_jagged_cat_jagged_out": {
-        "CUDA": dense_jagged_cat_jagged_out,
-    },
     "sll_jagged_self_substraction_jagged_out": {
         "CUDA": triton_jagged_self_substraction_jagged_out,
     },
-    "sll_jagged2_to_padded_dense": {
-        "CUDA": jagged2_to_padded_dense,
-        "AutogradCUDA": jagged2_to_padded_dense,
-    },
     "sll_jagged_dense_elementwise_mul_jagged_out": {
         "CUDA": jagged_dense_elementwise_mul_jagged_out,
         "AutogradCUDA": jagged_dense_elementwise_mul_jagged_out,
     },
-    "sll_array_jagged_bmm_jagged_out": {
-        "CUDA": array_jagged_bmm_jagged_out,
-        "AutogradCUDA": array_jagged_bmm_jagged_out,
-    },
-    "sll_jagged_jagged_bmm_jagged_out": {
-        "CUDA": jagged_jagged_bmm_jagged_out,
-        "AutogradCUDA": jagged_jagged_bmm_jagged_out,
-    },
 }
 
 for op_name, dispatches in sll_cpu_registrations.items():

fbgemm_gpu/fbgemm_gpu/sll/triton/__init__.py

@@ -7,13 +7,31 @@
 
 # pyre-strict
 
+from fbgemm_gpu.sll.triton.triton_dense_jagged_cat_jagged_out import (
+    dense_jagged_cat_jagged_out,
+)
+
+from fbgemm_gpu.sll.triton.triton_jagged2_to_padded_dense import (  # noqa F401
+    jagged2_to_padded_dense,
+    Jagged2ToPaddedDense,  # noqa F401
+)
+
 from fbgemm_gpu.sll.triton.triton_jagged_bmm import (  # noqa F401
     jagged_dense_bmm,
     jagged_jagged_bmm,
     JaggedDenseBmm,  # noqa F401
     JaggedJaggedBmm,  # noqa F401
 )
 
+from fbgemm_gpu.sll.triton.triton_jagged_bmm_jagged_out import (  # noqa F401
+    array_jagged_bmm_jagged_out,
+    ArrayJaggedBmmNopadding,  # noqa F401
+    jagged_jagged_bmm_jagged_out,
+    JaggedJaggedBmmNoPadding,  # noqa F401
+    triton_array_jagged_bmm_jagged_out,  # noqa F401
+    triton_jagged_jagged_bmm_jagged_out,  # noqa F401
+)
+
 from fbgemm_gpu.sll.triton.triton_jagged_dense_elementwise_add import (  # noqa F401
     jagged_dense_elementwise_add,
     JaggedDenseAdd,  # noqa F401
@@ -43,6 +61,9 @@
 
 # pyre-ignore[5]
 op_registrations = {
+    "sll_dense_jagged_cat_jagged_out": {
+        "CUDA": dense_jagged_cat_jagged_out,
+    },
     "sll_jagged_dense_bmm": {
         "CUDA": jagged_dense_bmm,
         "AutogradCUDA": jagged_dense_bmm,
@@ -51,6 +72,18 @@
         "CUDA": jagged_jagged_bmm,
         "AutogradCUDA": jagged_jagged_bmm,
     },
+    "sll_jagged2_to_padded_dense": {
+        "CUDA": jagged2_to_padded_dense,
+        "AutogradCUDA": jagged2_to_padded_dense,
+    },
+    "sll_array_jagged_bmm_jagged_out": {
+        "CUDA": array_jagged_bmm_jagged_out,
+        "AutogradCUDA": array_jagged_bmm_jagged_out,
+    },
+    "sll_jagged_jagged_bmm_jagged_out": {
+        "CUDA": jagged_jagged_bmm_jagged_out,
+        "AutogradCUDA": jagged_jagged_bmm_jagged_out,
+    },
     "sll_jagged_softmax": {
         "CUDA": jagged_softmax,
         "AutogradCUDA": jagged_softmax,

fbgemm_gpu/fbgemm_gpu/sll/triton/triton_dense_jagged_cat_jagged_out.py

@@ -0,0 +1,72 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-unsafe
+
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def dense_jagged_cat_jagged_out_kernel(
+    a_ptr,  # dense
+    b_ptr,  # jagged
+    c_ptr,  # jagged
+    b_offsets_ptr,
+    c_offsets_ptr,
+    max_seq_len,
+    BLOCK_SIZE: tl.constexpr,
+):
+    pid_batch = tl.program_id(0)
+    b_start = tl.load(b_offsets_ptr + pid_batch)
+    b_end = tl.load(b_offsets_ptr + pid_batch + 1)
+    c_start = b_start + pid_batch
+    N = b_end - b_start
+    N = tl.minimum(N, max_seq_len)
+
+    a = tl.load(a_ptr + pid_batch)
+    tl.store(c_ptr + c_start, a)
+
+    offs_k = tl.arange(0, BLOCK_SIZE)
+    for k in range(0, N, BLOCK_SIZE):
+        b_offset = k + offs_k
+        b_ptrs = b_ptr + b_start + b_offset
+        b = tl.load(b_ptrs, mask=b_offset < N, other=0.0)
+        tl.store(c_ptr + c_start + 1 + b_offset, b, mask=b_offset < N)
+    tl.store(c_offsets_ptr + pid_batch, b_start + pid_batch)
+
+
+def dense_jagged_cat_jagged_out(
+    a: torch.Tensor,
+    b: torch.Tensor,
+    b_offsets: torch.Tensor,
+    max_seq_len: int,
+):
+    assert a.is_contiguous()
+    assert b.is_contiguous()
+    assert b_offsets.is_contiguous()
+    B = a.size(0)
+    BLOCK_SIZE = 128
+    c = torch.zeros(b.size(0) + a.size(0), dtype=a.dtype, device=a.device)
+    c_offsets = torch.empty(
+        b_offsets.size(0), dtype=b_offsets.dtype, device=b_offsets.device
+    )  # B + 1
+
+    dense_jagged_cat_jagged_out_kernel[(B,)](
+        a,
+        b,
+        c,
+        b_offsets,
+        c_offsets,
+        max_seq_len,
+        # pyre-fixme[6]: For 7th argument expected `constexpr` but got `int`.
+        BLOCK_SIZE,
+    )
+
+    c_offsets[-1] = b_offsets[-1] + B
+
+    return c, c_offsets

fbgemm_gpu/fbgemm_gpu/sll/triton/triton_jagged2_to_padded_dense.py

@@ -0,0 +1,222 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-unsafe
+
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from .common import expect_contiguous
+
+
+@triton.jit
+def jagged2_to_padded_dense_kernel(
+    x_ptr,
+    lengths_ptr,
+    offsets_ptr,
+    output_dense_ptr,
+    stride_b,
+    stride_m,
+    stride_n,
+    max_length,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+):
+    pid_batch = tl.program_id(2)
+    pid_m = tl.program_id(0)
+    pid_n = tl.program_id(1)
+
+    begin = tl.load(offsets_ptr + pid_batch)
+    seqlen = tl.load(lengths_ptr + pid_batch)
+
+    seqlen = tl.minimum(seqlen, max_length)
+    if seqlen == 0:
+        return
+
+    offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+
+    x_ptrs = x_ptr + begin + offs_m[:, None] * seqlen + offs_n[None, :]
+    x = tl.load(x_ptrs, mask=((offs_m[:, None] < seqlen) & (offs_n[None, :] < seqlen)))
+
+    out_ptrs = (
+        output_dense_ptr
+        + pid_batch * stride_b
+        + offs_m[:, None] * stride_m
+        + offs_n[None, :] * stride_n
+    )
+    tl.store(
+        out_ptrs, x, mask=((offs_m[:, None] < seqlen) & (offs_n[None, :] < seqlen))
+    )
+
+
+@triton.jit
+def padded_dense_to_jagged2_kernel(
+    x_ptr,
+    lengths_ptr,
+    offsets_ptr,
+    output_jagged_ptr,
+    stride_b,
+    stride_m,
+    stride_n,
+    max_length,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+):
+    pid_batch = tl.program_id(2)
+    pid_m = tl.program_id(0)
+    pid_n = tl.program_id(1)
+
+    begin = tl.load(offsets_ptr + pid_batch)
+    # end = tl.load(offsets_ptr + pid_batch + 1)
+    seqlen = tl.load(lengths_ptr + pid_batch)
+
+    seqlen = tl.minimum(seqlen, max_length)
+
+    if seqlen == 0:
+        return
+
+    offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+
+    x_ptrs = (
+        x_ptr
+        + pid_batch * stride_b
+        + offs_m[:, None] * stride_m
+        + offs_n[None, :] * stride_n
+    )
+    x = tl.load(x_ptrs, mask=((offs_m[:, None] < seqlen) & (offs_n[None, :] < seqlen)))
+    out_ptrs = output_jagged_ptr + begin + offs_m[:, None] * seqlen + offs_n[None, :]
+    tl.store(
+        out_ptrs, x, mask=((offs_m[:, None] < seqlen) & (offs_n[None, :] < seqlen))
+    )
+
+
+def jagged2_to_padded_dense_fwd(
+    values: torch.Tensor,
+    lengths: torch.Tensor,
+    offsets: torch.Tensor,
+    max_length: int,
+    padding_value: float,
+) -> torch.Tensor:
+    B = offsets.size(0) - 1
+
+    output_dense = torch.full(
+        (B, max_length, max_length),
+        padding_value,
+        dtype=values.dtype,
+        device=values.device,
+    )
+    BLOCK_M = 32
+    BLOCK_N = 32
+    num_blocks_m = triton.cdiv(max_length, BLOCK_M)
+    num_blocks_n = triton.cdiv(max_length, BLOCK_N)
+    grid = (num_blocks_m, num_blocks_n, B)
+
+    jagged2_to_padded_dense_kernel[grid](
+        values,
+        lengths,
+        offsets,
+        output_dense,
+        output_dense.stride(0),
+        output_dense.stride(1),
+        output_dense.stride(2),
+        max_length,
+        # pyre-fixme[6]: Incompatible parameter type [6]: expected `constexpr` but got `int`.
+        BLOCK_M,
+        # pyre-fixme[6]: Incompatible parameter type [6]: expected `constexpr` but got `int`.
+        BLOCK_N,
+    )
+
+    return output_dense
+
+
+def padded_dense_to_jagged2_fwd(
+    values: torch.Tensor,
+    lengths: torch.Tensor,
+    offsets: torch.Tensor,
+    max_length: int,
+) -> torch.Tensor:
+    B = values.size(0)
+    output_jagged = torch.empty(
+        int(offsets[-1]), dtype=values.dtype, device=values.device
+    )
+    BLOCK_M = 32
+    BLOCK_N = 32
+    num_blocks_m = triton.cdiv(max_length, BLOCK_M)
+    num_blocks_n = triton.cdiv(max_length, BLOCK_N)
+    grid = (num_blocks_m, num_blocks_n, B)
+
+    padded_dense_to_jagged2_kernel[grid](
+        values,
+        lengths,
+        offsets,
+        output_jagged,
+        values.stride(0),
+        values.stride(1),
+        values.stride(2),
+        max_length,
+        # pyre-fixme[6]: Incompatible parameter type [6]: expected `constexpr` but got `int`.
+        BLOCK_M,
+        # pyre-fixme[6]: Incompatible parameter type [6]: expected `constexpr` but got `int`.
+        BLOCK_N,
+    )
+
+    return output_jagged
+
+
+class Jagged2ToPaddedDense(torch.autograd.Function):
+    @staticmethod
+    # pyre-fixme
+    def forward(
+        ctx,
+        values: torch.Tensor,
+        offsets: torch.Tensor,
+        max_length: int,
+        padding_value: float,
+    ) -> torch.Tensor:
+        lengths_square = offsets[1:] - offsets[0:-1:1]
+        lengths = torch.sqrt(lengths_square).to(torch.int32)
+
+        ctx.max_length = max_length
+        ctx.save_for_backward(lengths, offsets)
+
+        output = jagged2_to_padded_dense_fwd(
+            values, lengths, offsets, max_length, padding_value
+        )
+        return output
+
+    @staticmethod
+    # pyre-fixme
+    def backward(
+        ctx, grad_output: torch.Tensor
+    ) -> Tuple[torch.Tensor, None, None, None]:
+        max_length = ctx.max_length
+        (lengths, offsets) = ctx.saved_tensors
+        grad_in = padded_dense_to_jagged2_fwd(grad_output, lengths, offsets, max_length)
+        return (grad_in, None, None, None)
+
+
+def jagged2_to_padded_dense(
+    values: torch.Tensor,
+    offsets: torch.Tensor,
+    max_length: int,
+    padding_value: float = 0.0,
+) -> torch.Tensor:
+    """
+    values: jagged tensor with size [sum(Ni * Ni)]
+    offsets: offsets for jagged tensor, with size [B + 1]
+    max_length: maximum sequence length in the batch
+    padding_value: value to use for padding
+    return padded dense tensor of size [B, N, N]
+    """
+    values = expect_contiguous(values)
+    offsets = expect_contiguous(offsets)
+
+    return Jagged2ToPaddedDense.apply(values, offsets, max_length, padding_value)