File size: 13,927 Bytes
8c8128e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
# Copyright (c) 2025, Wentao Guo, Ted Zadouri, Tri Dao.


import cutlass
import cutlass.cute as cute

from cutlass import Int32, const_expr


def transpose_view(a: cute.Tensor) -> cute.Tensor:
    """Transpose the first two dimensions of a tensor on smem."""
    shape = (a.shape[1], a.shape[0], *a.shape[2:])
    order = (1, 0, *range(2, cute.rank(a)))
    return cute.composition(a, cute.make_ordered_layout(shape, order=order))


def select(a: cute.Tensor, mode: list[int]) -> cute.Tensor:
    return cute.make_tensor(a.iterator, cute.select(a.layout, mode))


def expand(a: cute.Tensor, dim: int, size: Int32 | int) -> cute.Tensor:
    shape = (*a.shape[:dim], size, *a.shape[dim:])
    stride = (*a.layout.stride[:dim], 0, *a.layout.stride[dim:])
    return cute.make_tensor(a.iterator, cute.make_layout(shape, stride=stride))


@cute.jit
def permute_gated_Cregs_b16(t: cute.Tensor) -> None:
    assert t.element_type.width == 16
    assert cute.size(t.shape) % 4 == 0, "Tensor size must be a multiple of 4 for b16 permutation"
    t_u32 = cute.recast_tensor(t, Int32)

    quad_idx = cute.arch.lane_idx() % 4
    lane_03 = quad_idx == 0 or quad_idx == 3
    selector_upper = Int32(0x5410) if lane_03 else Int32(0x1054)
    selector_lower = Int32(0x7632) if lane_03 else Int32(0x3276)
    # upper_map = [0, 3, 1, 2]
    # lower_map = [1, 2, 0, 3]
    # upper_idx = upper_map[quad_idx]
    # indexing isn't supported so we have to do arithmetic
    upper_idx = quad_idx // 2 if quad_idx % 2 == 0 else 3 - quad_idx // 2
    lower_idx = upper_idx ^ 1

    # 1 -> 0b11111, 2 -> 0b11110, 4 -> 0b11100, 8 -> 0b11000, 16 -> 0b10000, 32 -> 0b00000
    width = 4
    mask = cute.arch.WARP_SIZE - width
    clamp = cute.arch.WARP_SIZE - 1
    mask_and_clamp = mask << 8 | clamp

    for i in cutlass.range(cute.size(t_u32.shape) // 2, unroll_full=True):
        upper, lower = t_u32[i * 2 + 0], t_u32[i * 2 + 1]
        upper0 = upper if lane_03 else lower
        lower0 = lower if lane_03 else upper
        upper0 = cute.arch.shuffle_sync(upper0, offset=upper_idx, mask_and_clamp=mask_and_clamp)
        lower0 = cute.arch.shuffle_sync(lower0, offset=lower_idx, mask_and_clamp=mask_and_clamp)
        t_u32[i * 2 + 0] = cute.arch.prmt(upper0, lower0, selector_upper)
        t_u32[i * 2 + 1] = cute.arch.prmt(upper0, lower0, selector_lower)


@cute.jit
def permute_Cregs_b32_for_stsm(t: cute.Tensor) -> None:
    """Permute and shuffle within 4 threads to change the layout from
     T0 | T1  | T2  | T3
    a b | c d | e f | g h
    to
    T0 | T1 | T2 | T3 | T0 | T1 | T2 | T3
    a  | b  | c  | d  | e  | f  | g  | h
    This is so that we can use STSM (instead of STS.64) to store C registers without bank conflict.
    """

    assert t.element_type.width == 32
    assert cute.size(t.shape) % 4 == 0, "Tensor size must be a multiple of 4 for b32 permutation"

    quad_idx = cute.arch.lane_idx() % 4
    # left_map = [0, 2, 1, 3]
    # right_map = [2, 0, 3, 1]
    # indexing isn't supported so we have to do arithmetic
    left_idx = quad_idx // 2 if quad_idx % 2 == 0 else 2 + quad_idx // 2
    right_idx = left_idx ^ 0b10

    # 1 -> 0b11111, 2 -> 0b11110, 4 -> 0b11100, 8 -> 0b11000, 16 -> 0b10000, 32 -> 0b00000
    width = 4
    mask = cute.arch.WARP_SIZE - width
    clamp = cute.arch.WARP_SIZE - 1
    mask_and_clamp = mask << 8 | clamp

    for i in cutlass.range(cute.size(t.shape) // 4, unroll_full=True):
        for r in cutlass.range(2, unroll_full=True):
            left, right = t[i * 4 + r * 2 + 0], t[i * 4 + r * 2 + 1]
            # a b | c d | e f | g h -> a b | c d | f e | h g
            left0 = left if quad_idx < 2 else right
            right0 = right if quad_idx < 2 else left
            # a b | c d | f e | h g -> a b | f d | c e | h g
            left0 = cute.arch.shuffle_sync(left0, offset=left_idx, mask_and_clamp=mask_and_clamp)
            # a b | f d | c e | h g -> a e | f b | c g | h d
            right0 = cute.arch.shuffle_sync(right0, offset=right_idx, mask_and_clamp=mask_and_clamp)
            # a e | f b | c g | h d -> a e | b f | c g | d h
            t[i * 4 + r * 2 + 0] = left0 if quad_idx % 2 == 0 else right0
            t[i * 4 + r * 2 + 1] = right0 if quad_idx % 2 == 0 else left0
        t[i * 4 + 1], t[i * 4 + 2] = t[i * 4 + 2], t[i * 4 + 1]


@cute.jit
def permute_Cregs_b32_for_ldsm(t: cute.Tensor) -> None:
    """Permute and shuffle within 4 threads to change the layout from
    T0 | T1 | T2 | T3 | T0 | T1 | T2 | T3
    a  | b  | c  | d  | e  | f  | g  | h
    to
     T0 | T1  | T2  | T3
    a b | c d | e f | g h
    This is so that we can use LDSM (instead of LDS.64) to store C registers without bank conflict.
    """

    assert t.element_type.width == 32
    assert cute.size(t.shape) % 4 == 0, "Tensor size must be a multiple of 4 for b32 permutation"

    quad_idx = cute.arch.lane_idx() % 4
    # left_map = [0, 2, 1, 3]
    # right_map = [1, 3, 0, 2]
    # indexing isn't supported so we have to do arithmetic
    left_idx = quad_idx // 2 if quad_idx % 2 == 0 else 2 + quad_idx // 2
    right_idx = left_idx ^ 0b01

    # 1 -> 0b11111, 2 -> 0b11110, 4 -> 0b11100, 8 -> 0b11000, 16 -> 0b10000, 32 -> 0b00000
    width = 4
    mask = cute.arch.WARP_SIZE - width
    clamp = cute.arch.WARP_SIZE - 1
    mask_and_clamp = mask << 8 | clamp

    # This is just the inverse of permute_Cregs_b32_for_stsm
    for i in cutlass.range(cute.size(t.shape) // 4, unroll_full=True):
        t[i * 4 + 1], t[i * 4 + 2] = t[i * 4 + 2], t[i * 4 + 1]
        for r in cutlass.range(2, unroll_full=True):
            left, right = t[i * 4 + r * 2 + 0], t[i * 4 + r * 2 + 1]
            # a e | b f | c g | d h -> a e | f b | c g | h d
            left0 = left if quad_idx % 2 == 0 else right
            right0 = right if quad_idx % 2 == 0 else left
            # a e | f b | c g | h d -> a b | f d | c e | h g
            right0 = cute.arch.shuffle_sync(right0, offset=right_idx, mask_and_clamp=mask_and_clamp)
            # a b | f d | c e | h g -> a b | c d | f e | h g
            left0 = cute.arch.shuffle_sync(left0, offset=left_idx, mask_and_clamp=mask_and_clamp)
            # a b | c d | f e | h g -> a b | c d | e f | g h
            t[i * 4 + r * 2 + 0] = left0 if quad_idx < 2 else right0
            t[i * 4 + r * 2 + 1] = right0 if quad_idx < 2 else left0


@cute.jit
def concat_layout(*layouts: cute.Layout) -> cute.Layout:
    return cute.make_layout(
        tuple(l.shape for l in layouts),
        stride=tuple(l.stride for l in layouts),
    )


def convert_layout_acc_mn(acc_layout: cute.Layout, transpose: bool = False) -> cute.Layout:
    """
    For Sm80, convert ((2, 2), MMA_M, MMA_N, ...) to ((2, MMA_M), (2, MMA_N), ...).
    For Sm90, convert ((2, 2, V), MMA_M, MMA_N, ...) to ((2, MMA_M), (2, V, MMA_N), ...).
    """
    acc_layout_col_major = cute.make_layout(acc_layout.shape)
    shape = (
        (acc_layout_col_major.shape[0][1], acc_layout_col_major.shape[1]),  # MMA_M
        (
            acc_layout_col_major.shape[0][0],
            *acc_layout_col_major.shape[0][2:],
            acc_layout_col_major.shape[2],
        ),  # MMA_N
        *acc_layout_col_major.shape[3:],
    )
    stride = (
        (acc_layout_col_major.stride[0][1], acc_layout_col_major.stride[1]),  # MMA_M
        (
            acc_layout_col_major.stride[0][0],
            *acc_layout_col_major.stride[0][2:],
            acc_layout_col_major.stride[2],
        ),  # MMA_N
        *acc_layout_col_major.stride[3:],
    )
    if const_expr(transpose):
        shape = (shape[1], shape[0], *shape[2:])
        stride = (stride[1], stride[0], *stride[2:])
    acc_layout_mn = cute.make_layout(shape, stride=stride)
    return cute.composition(acc_layout, acc_layout_mn)


def make_acc_tensor_mn_view(acc: cute.Tensor, transpose: bool = False) -> cute.Tensor:
    return cute.make_tensor(acc.iterator, convert_layout_acc_mn(acc.layout, transpose=transpose))


def reshape_acc_to_mn(acc: cute.Tensor, transpose: bool = False) -> cute.Tensor:
    return cute.make_tensor(acc.iterator, convert_layout_acc_mn(acc.layout, transpose=transpose))


@cute.jit
def convert_layout_acc_frgA(acc_layout: cute.Layout) -> cute.Layout:
    # For back to back gemm, convert layout of acc0 to gemm 1 accept layout.
    # For Sm80, as the mma instruction shape is 16x8x16, we need to convert from (4, MMA_M, MMA_N) to ((4, 2), MMA_M, MMA_N / 2)
    # For Sm90, FP16/BF16, convert acc_layout from ((2, 2, N / 8), MMA_M, MMA_N) to ((2, 2, 2), MMA_M, (N / 16, MMA_N))
    # If N / 8 is odd, we'll convert to ((2, 2, 1), MMA_M, N / 8, MMA_N).
    # TODO: Sm90 FP8
    if const_expr(cute.rank(acc_layout.shape[0]) == 3):  # Sm90
        div = 2 if const_expr(acc_layout.shape[0][2] % 2 == 0) else 1
        l = cute.logical_divide(
            acc_layout, ((None, None, div), None, None)
        )  # ((2, 2, (2, N / 16)), MMA_M, MMA_N)
        rA_mma_view = cute.make_layout(
            (
                (l.shape[0][0], l.shape[0][1], l.shape[0][2][0]),
                l.shape[1],
                (l.shape[0][2][1], l.shape[2]),
            ),
            stride=(
                (l.stride[0][0], l.stride[0][1], l.stride[0][2][0]),
                l.stride[1],
                (l.stride[0][2][1], l.stride[2]),
            ),
        )
    else:  # Sm80
        # (4, MMA_M, MMA_N) -> (4, MMA_M, (2, MMA_N / 2))
        l = cute.logical_divide(acc_layout, (None, None, 2))
        rA_mma_view = cute.make_layout(
            (
                (l.shape[0], l.shape[2][0]),
                l.shape[1],
                l.shape[2][1],
            ),
            stride=(
                (l.stride[0], l.stride[2][0]),
                l.stride[1],
                l.stride[2][1],
            ),
        )
    return rA_mma_view


def reshape_acc_to_frgA(acc: cute.Tensor) -> cute.Tensor:
    return cute.make_tensor(acc.iterator, convert_layout_acc_frgA(acc.layout))


def convert_layout_zero_stride(
    input: cute.Tensor | cute.Layout, ref_layout: cute.Layout
) -> cute.Layout:
    layout = input.layout if const_expr(isinstance(input, cute.Tensor)) else input
    # Group the modes with non-zero stride in the ref_layout together,
    # and the modes with zero stride together
    layout_flat = cute.flatten(layout)
    ref_layout_flat = cute.flatten(ref_layout)
    nonzero_modes = [i for i in range(cute.rank(layout_flat)) if ref_layout_flat[i].stride != 0]
    zero_modes = [i for i in range(cute.rank(layout_flat)) if ref_layout_flat[i].stride == 0]
    # There's an edge case when all modes are zero stride
    new_shape = (
        tuple(layout_flat[i].shape for i in nonzero_modes) if len(nonzero_modes) > 0 else (1,),
        tuple(layout_flat[i].shape for i in zero_modes),
    )
    new_stride = (
        tuple(layout_flat[i].stride for i in nonzero_modes) if len(nonzero_modes) > 0 else (0,),
        tuple(layout_flat[i].stride for i in zero_modes),
    )
    out_layout = cute.make_layout(new_shape, stride=new_stride)
    if const_expr(isinstance(input, cute.Tensor)):
        return cute.make_tensor(input.iterator, out_layout)
    else:
        return out_layout


def mma_partition_C_vec(
    sVec: cute.Tensor, thr_mma: cute.core.ThrMma, expand_shape: int, is_colvec: bool
) -> cute.Tensor:
    assert cute.rank(sVec) == 2
    assert sVec.stride[0] == 1
    stage = sVec.shape[1]
    shape = (
        (sVec.shape[0], expand_shape, stage)
        if const_expr(is_colvec)
        else (expand_shape, sVec.shape[0], stage)
    )
    stride = (1, 0, sVec.stride[1]) if const_expr(is_colvec) else (0, 1, sVec.stride[1])
    sVec_mma = cute.make_tensor(sVec.iterator, cute.make_layout(shape, stride=stride))
    tC_sVec = make_acc_tensor_mn_view(thr_mma.partition_C(sVec_mma))
    return tC_sVec[None, 0, None] if const_expr(is_colvec) else tC_sVec[0, None, None]


def mma_partition_A_vec(
    sVec: cute.Tensor, thr_mma: cute.core.ThrMma, expand_shape: int, is_colvec: bool
) -> cute.Tensor:
    assert cute.rank(sVec) == 2
    assert sVec.stride[0] == 1
    stage = sVec.shape[1]
    shape = (
        (sVec.shape[0], expand_shape, stage)
        if const_expr(is_colvec)
        else (expand_shape, sVec.shape[0], stage)
    )
    stride = (1, 0, sVec.stride[1]) if const_expr(is_colvec) else (0, 1, sVec.stride[1])
    sVec_mma = cute.make_tensor(sVec.iterator, cute.make_layout(shape, stride=stride))
    tC_sVec = make_acc_tensor_mn_view(thr_mma.partition_A(sVec_mma))
    return tC_sVec[None, 0, None] if const_expr(is_colvec) else tC_sVec[0, None, None]


def copy_partition_S_vec(
    sVec: cute.Tensor, thr_copy: cute.core.ThrCopy, expand_shape: int, is_colvec: bool
) -> cute.Tensor:
    assert cute.rank(sVec) == 2
    assert sVec.stride[0] == 1
    stage = sVec.shape[1]
    shape = (
        (sVec.shape[0], expand_shape, stage)
        if const_expr(is_colvec)
        else (expand_shape, sVec.shape[0], stage)
    )
    stride = (1, 0, sVec.stride[1]) if const_expr(is_colvec) else (0, 1, sVec.stride[1])
    sVec_thr = cute.make_tensor(sVec.iterator, cute.make_layout(shape, stride=stride))
    tC_sVec = reshape_acc_to_mn(thr_copy.partition_S(sVec_thr))
    return tC_sVec[None, 0, None] if const_expr(is_colvec) else tC_sVec[0, None, None]


def copy_partition_D_vec(
    sVec: cute.Tensor, thr_copy: cute.core.ThrCopy, expand_shape: int, is_colvec: bool
) -> cute.Tensor:
    assert cute.rank(sVec) == 2
    assert sVec.stride[0] == 1
    stage = sVec.shape[1]
    shape = (
        (sVec.shape[0], expand_shape, stage)
        if const_expr(is_colvec)
        else (expand_shape, sVec.shape[0], stage)
    )
    stride = (1, 0, sVec.stride[1]) if const_expr(is_colvec) else (0, 1, sVec.stride[1])
    sVec_thr = cute.make_tensor(sVec.iterator, cute.make_layout(shape, stride=stride))
    tC_sVec = reshape_acc_to_mn(thr_copy.partition_D(sVec_thr))
    return tC_sVec[None, 0, None] if const_expr(is_colvec) else tC_sVec[0, None, None]