triton.language.make_block_ptr

1. OP 概述

简介：创建指向GM上张量的指针原型：

triton.language.make_block_ptr(
 base: triton.PointerType,
 shape: List[tensor],
 strides: tuple(int | constexpr),
 offsets: tuple(int | constexpr),
 block_shape:tuple(int | constexpr),
 order:tuple(constexpr),
 _semantic=None
)

2. OP 规格

2.1 参数说明

参数名	类型	说明
`base`	`triton.PointerType`	张量的基指针
`shape`	`tuple(int \| constexpr)`	张量在GM上的形状
`strides`	`tuple(int \| constexpr)`	张量各维度的步长列表
`offsets`	`tuple(int \| constexpr)`	张量各维度的基址偏移量列表
`block_shape`	`tuple(constexpr)`	单次从全局内存加载 / 存储的块的形状
`order`	`tuple(constexpr)`	单次从全局内存加载 / 存储的块的形状
`_semantic`	-	保留参数，暂不支持外部调用

返回值：pointer_type<blocked<shape, element_type>>：指向tensor的指针

2.2 支持规格

2.2.1 DataType 支持

	int8	int16	int32	uint8	uint16	uint32	uint64	int64	fp16	fp32	fp64	bf16	bool
GPU	√	√	√	√	√	√	√	√	√	√	√	√	×
Ascend A2/A3	√	√	√	×	×	×	×	√	√	√	×	√	×

结论：Ascend 对比 GPU 缺失uint8、uint16、uint32、uint64、fp64的支持能力（硬件限制）。

2.2.2 Shape 支持

	支持维度范围
GPU	仅支持 1~5维 tensor
Ascend A2/A3	仅支持 1~5维 tensor

结论：在 Shape 方面，GPU 与 Ascend 平台无差异，均支持 1 至 5 维张量。

2.2.3 社区约束

tl.make_block_ptr的结果不允许进行算数运算，需要改变偏移量时，可以通过：

重新调用make_block_ptr，修改offset参数:

for block_idx in range(pid, NUM_BLOCKS, 20):
    task_hz_idx = block_idx // NUM_BLOCKS_M
    task_m_idx = block_idx % NUM_BLOCKS_M
    off_z = task_hz_idx // H
    off_h = task_hz_idx % H
    qvk_offset = off_z.to(tl.int64) * stride_qz + off_h.to(tl.int64) * stride_qh
    # Create block pointers for Q, K, V, Output
    Q_block_ptr = tl.make_block_ptr(
        base=Q + qvk_offset,
        shape=(N_CTX, HEAD_DIM),
        strides=(stride_qm, stride_qk),
        offsets=(task_m_idx * BLOCK_M, 0),
        block_shape=(BLOCK_M, HEAD_DIM),
        order=(1, 0),
    )

调用tl.advance调整偏移量:

block_ptr_in=tl.make_block_ptr(
    base = x_ptr,
    shape = (XB,YB,ZB),
    strides = (YB*ZB,ZB,1),
    offsets = (9,6,5),
    block_shape = (XB,YB,ZB),
    order = (2,1,0),
)
bbptr = tl.advance(block_ptr_in,(-9,-6,-5))

2.3 特殊限制说明

相对社区能力缺失且无法实现

Ascend 对比 GPU 缺失uint8、uint16、uint32、uint64、fp64的支持能力（硬件限制）。
Ascend只允许通过调整order参数的顺序来表达转置语义，不能通过调整stride参数的顺序实现转置语义。

差异点	描述	解决途径
与分支、循环语句搭配使用时的泛化性问题	当前`tl.make_tensor_ptr`，如果与较复杂的循环和分支语句搭配使用，可能会出现编译问题	大量泛化测试暴露问题，迭代解决

2.4 使用方法

以下示例实现了通过tl.make_block_ptr读取一个tensor并将其转置的功能：

import torch
import torch_npu
import triton
import triton.language as tl
import pytest
import test_common

@pytest.mark.parametrize('shape', [(1, 4, 2)])
@pytest.mark.parametrize('permute_order', [(2, 0, 1)])
def test_makeblockptr_order(shape, permute_order):

    @triton.jit
    def triton_kernel(in0_ptr: tl.tensor, # of tl.pointer_type
                out0_ptr: tl.tensor, # of tl.pointer_type
                in0_stride0: int, in0_stride1: int, in0_stride2: int, # strides for in0
                in0_stride_order0: tl.constexpr, in0_stride_order1: tl.constexpr, in0_stride_order2: tl.constexpr, # stride order for in0
                out0_stride0: int, out0_stride1: int, out0_stride2: int, # strides for out0
                out0_stride_order0: tl.constexpr, out0_stride_order1: tl.constexpr, out0_stride_order2: tl.constexpr, # stride order for out0
                s0: int, s1: int, s2: int,
                tile_size0: tl.constexpr, tile_size1: tl.constexpr, tile_size2: tl.constexpr,
                ):
        tile_id0 = tl.program_id(axis=0)
        tile_id1 = tl.program_id(axis=1)
        tile_id2 = tl.program_id(axis=2)
        offset0 = (tile_id0 * tile_size0).to(tl.int32)
        offset1 = (tile_id1 * tile_size1).to(tl.int32)
        offset2 = (tile_id2 * tile_size2).to(tl.int32)
        in0_bptr = tl.make_block_ptr(in0_ptr,
                                    (s0, s1, s2),
                                    (in0_stride0, in0_stride1, in0_stride2),
                                    (offset0, offset1, offset2),
                                    (tile_size0, tile_size1, tile_size2),
                                    order=(in0_stride_order0, in0_stride_order1, in0_stride_order2))
        in0 = tl.load(in0_bptr, boundary_check=(in0_stride_order0, in0_stride_order1, in0_stride_order2)).to(in0_ptr.type.element_ty)

        out0 = in0

        out0_bptr = tl.make_block_ptr(out0_ptr, (s0, s1, s2), (out0_stride0, out0_stride1, out0_stride2), (offset0, offset1, offset2), (tile_size0, tile_size1, tile_size2),
                                    order=(out0_stride_order0, out0_stride_order1, out0_stride_order2))
        tl.store(out0_bptr, out0.to(out0_bptr.type.element_ty), boundary_check=(out0_stride_order0, out0_stride_order1, out0_stride_order2))

    def triton_func(in0: torch.Tensor, permute_order):
        # in fact, it adjusts the layout metadata instead of doing a real permutation.
        in0_permuted_tmp = in0.permute(permute_order)
        in0_permuted_shape = in0_permuted_tmp.size()
        in0_permuted_strides = in0_permuted_tmp.stride()
        in0_stride_order = [len(permute_order)-1-i for i in permute_order]
        shape = (in0_permuted_shape[0], in0_permuted_shape[1], in0_permuted_shape[2])
        tile_sizes = (shape[0], shape[1], shape[2])
        out0 = torch.empty(shape, dtype=in0.dtype, device=in0.device)
        out0_strides = out0.stride()
        out0_stride_order = [len(permute_order)-1-i for i in range(len(permute_order))]
        grid = (shape[0]//tile_sizes[0], shape[1]//tile_sizes[1], shape[2]//tile_sizes[2])
        triton_kernel[grid](
                in0, out0,
                in0_permuted_strides[0], in0_permuted_strides[1], in0_permuted_strides[2], # stride for in0
                in0_stride_order[0], in0_stride_order[1], in0_stride_order[2], # stride order for in0
                out0_strides[0], out0_strides[1], out0_strides[2], # stride for out0
                out0_stride_order[0], out0_stride_order[1], out0_stride_order[2], # stride orderfor out0
                shape[0], shape[1], shape[2], # task indexing space
                tile_size0=tile_sizes[0],
                tile_size1=tile_sizes[1],
                tile_size2=tile_sizes[2],
            )
        return out0

    x0 = torch.randint(0, 9, shape, dtype=torch.int32).npu()
    torch_ref = torch.permute(x0, permute_order)
    triton_cal = triton_func(x0, permute_order)
    test_common.validate_cmp("int32", triton_cal, torch_ref)