Skip to content

Convert snitch stream to snitch

convert_snitch_stream_to_snitch

LowerStreamingRegionOp

Bases: RewritePattern

Source code in xdsl/backend/riscv/lowering/convert_snitch_stream_to_snitch.py 
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
class LowerStreamingRegionOp(RewritePattern):
    @op_type_rewrite_pattern
    def match_and_rewrite(
        self, op: snitch_stream.StreamingRegionOp, rewriter: PatternRewriter, /
    ):
        # Set up stream stride patterns
        # Set up stream pointers
        # Insert stream begin
        # Inline body
        # Insert stream end
        input_count = len(op.inputs)
        output_count = len(op.outputs)
        stream_count = input_count + output_count

        # If there is a single pattern specified, then it should be set for all streams
        if len(op.stride_patterns) == 1 and stream_count != 1:
            pattern = op.stride_patterns.data[0]
            patterns = (pattern,) * stream_count
            # Set same pattern for all data movers
            insert_stride_pattern_ops(
                rewriter,
                op,
                pattern.ub,
                pattern.strides,
                pattern.repeat,
                builtin.IntAttr(31),
            )
        else:
            patterns = op.stride_patterns.data
            # Set separate pattern per data mover
            for dm, pattern in enumerate(patterns):
                insert_stride_pattern_ops(
                    rewriter,
                    op,
                    pattern.ub,
                    pattern.strides,
                    pattern.repeat,
                    builtin.IntAttr(dm),
                )

        dms = tuple(range(stream_count))

        set_source_ops = tuple(
            snitch.SsrSetDimensionSourceOp(
                input,
                dm=builtin.IntAttr(dm),
                dimension=builtin.IntAttr(pattern.rank() - 1),
            )
            for input, pattern, dm in zip(
                op.inputs, patterns[:input_count], dms[:input_count], strict=True
            )
        )

        rewriter.insert_op(set_source_ops)

        set_destination_ops = tuple(
            snitch.SsrSetDimensionDestinationOp(
                output,
                dm=builtin.IntAttr(dm),
                dimension=builtin.IntAttr(pattern.rank() - 1),
            )
            for output, pattern, dm in zip(
                op.outputs, patterns[input_count:], dms[input_count:], strict=True
            )
        )
        rewriter.insert_op(set_destination_ops)

        block = op.body.block

        rewriter.insert_op(enable_op := snitch.SsrEnableOp(block.arg_types))

        for val, arg in zip(enable_op.streams, block.args):
            arg.replace_all_uses_with(val)

        for arg in reversed(block.args):
            rewriter.erase_block_argument(arg)

        rewriter.inline_block(block, InsertPoint.before(op))

        rewriter.replace_op(op, snitch.SsrDisableOp())

match_and_rewrite(op: snitch_stream.StreamingRegionOp, rewriter: PatternRewriter)

Source code in xdsl/backend/riscv/lowering/convert_snitch_stream_to_snitch.py 
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
@op_type_rewrite_pattern
def match_and_rewrite(
    self, op: snitch_stream.StreamingRegionOp, rewriter: PatternRewriter, /
):
    # Set up stream stride patterns
    # Set up stream pointers
    # Insert stream begin
    # Inline body
    # Insert stream end
    input_count = len(op.inputs)
    output_count = len(op.outputs)
    stream_count = input_count + output_count

    # If there is a single pattern specified, then it should be set for all streams
    if len(op.stride_patterns) == 1 and stream_count != 1:
        pattern = op.stride_patterns.data[0]
        patterns = (pattern,) * stream_count
        # Set same pattern for all data movers
        insert_stride_pattern_ops(
            rewriter,
            op,
            pattern.ub,
            pattern.strides,
            pattern.repeat,
            builtin.IntAttr(31),
        )
    else:
        patterns = op.stride_patterns.data
        # Set separate pattern per data mover
        for dm, pattern in enumerate(patterns):
            insert_stride_pattern_ops(
                rewriter,
                op,
                pattern.ub,
                pattern.strides,
                pattern.repeat,
                builtin.IntAttr(dm),
            )

    dms = tuple(range(stream_count))

    set_source_ops = tuple(
        snitch.SsrSetDimensionSourceOp(
            input,
            dm=builtin.IntAttr(dm),
            dimension=builtin.IntAttr(pattern.rank() - 1),
        )
        for input, pattern, dm in zip(
            op.inputs, patterns[:input_count], dms[:input_count], strict=True
        )
    )

    rewriter.insert_op(set_source_ops)

    set_destination_ops = tuple(
        snitch.SsrSetDimensionDestinationOp(
            output,
            dm=builtin.IntAttr(dm),
            dimension=builtin.IntAttr(pattern.rank() - 1),
        )
        for output, pattern, dm in zip(
            op.outputs, patterns[input_count:], dms[input_count:], strict=True
        )
    )
    rewriter.insert_op(set_destination_ops)

    block = op.body.block

    rewriter.insert_op(enable_op := snitch.SsrEnableOp(block.arg_types))

    for val, arg in zip(enable_op.streams, block.args):
        arg.replace_all_uses_with(val)

    for arg in reversed(block.args):
        rewriter.erase_block_argument(arg)

    rewriter.inline_block(block, InsertPoint.before(op))

    rewriter.replace_op(op, snitch.SsrDisableOp())

ConvertSnitchStreamToSnitch dataclass

Bases: ModulePass

Source code in xdsl/backend/riscv/lowering/convert_snitch_stream_to_snitch.py 
189
190
191
192
193
194
195
196
197
198
class ConvertSnitchStreamToSnitch(ModulePass):
    name = "convert-snitch-stream-to-snitch"

    def apply(self, ctx: Context, op: builtin.ModuleOp) -> None:
        # StridedWrite and StridePattern ops are rewritten to remove their results, so we
        # have to first lower the ops that use the results in `stream`, and then the ops
        # themselves.
        PatternRewriteWalker(
            LowerStreamingRegionOp(), apply_recursively=False
        ).rewrite_module(op)

name = 'convert-snitch-stream-to-snitch' class-attribute instance-attribute

apply(ctx: Context, op: builtin.ModuleOp) -> None

Source code in xdsl/backend/riscv/lowering/convert_snitch_stream_to_snitch.py 
192
193
194
195
196
197
198
def apply(self, ctx: Context, op: builtin.ModuleOp) -> None:
    # StridedWrite and StridePattern ops are rewritten to remove their results, so we
    # have to first lower the ops that use the results in `stream`, and then the ops
    # themselves.
    PatternRewriteWalker(
        LowerStreamingRegionOp(), apply_recursively=False
    ).rewrite_module(op)

insert_stride_pattern_ops(rewriter: PatternRewriter, target_op: Operation, ub: builtin.ArrayAttr[builtin.IntAttr], strides: builtin.ArrayAttr[builtin.IntAttr], repeat: builtin.IntAttr, dm: builtin.IntAttr)

ub and strides must go from the outermost dimension inwards

Source code in xdsl/backend/riscv/lowering/convert_snitch_stream_to_snitch.py 
 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
def insert_stride_pattern_ops(
    rewriter: PatternRewriter,
    target_op: Operation,
    ub: builtin.ArrayAttr[builtin.IntAttr],
    strides: builtin.ArrayAttr[builtin.IntAttr],
    repeat: builtin.IntAttr,
    dm: builtin.IntAttr,
):
    """
    `ub` and `strides` must go from the outermost dimension inwards
    """
    # reference implementation:
    # https://github.com/pulp-platform/snitch/blob/d026f47843f0ea6c269244c4e6851e0e09141ec3/sw/snRuntime/src/ssr.h#L73
    #
    # 4d loop reproduced here:
    #
    # // Configure an SSR data mover for a 4D loop nest.
    # // b0: Inner-most bound (limit of loop)
    # // b3: Outer-most bound (limit of loop)
    # // s0: increment size of inner-most loop
    # inline void snrt_ssr_loop_4d(enum snrt_ssr_dm dm, size_t b0, size_t b1,
    #                              size_t b2, size_t b3, size_t s0, size_t s1,
    #                              size_t s2, size_t s3) {
    #     --b0;
    #     write_ssr_cfg(REG_BOUNDS + 0, dm, b0);
    #     --b1;
    #     write_ssr_cfg(REG_BOUNDS + 1, dm, b1);
    #     --b2;
    #     write_ssr_cfg(REG_BOUNDS + 2, dm, b2);
    #     --b3;
    #     write_ssr_cfg(REG_BOUNDS + 3, dm, b3);
    #     size_t a = 0;
    #     write_ssr_cfg(REG_STRIDES + 0, dm, s0 - a);
    #     a += s0 * b0;
    #     write_ssr_cfg(REG_STRIDES + 1, dm, s1 - a);
    #     a += s1 * b1;
    #     write_ssr_cfg(REG_STRIDES + 2, dm, s2 - a);
    #     a += s2 * b2;
    #     write_ssr_cfg(REG_STRIDES + 3, dm, s3 - a);
    #     a += s3 * b3;
    # }

    rank = len(ub)
    if rank > 4:
        raise NotImplementedError(
            "Only 1d, 2d, 3d, or 4d loop stride patterns are supported"
        )

    ints = tuple(builtin.IntAttr(i) for i in range(rank))

    b_ops = tuple(riscv.LiOp(b.data) for b in reversed(ub.data))
    new_b_ops = tuple(riscv.AddiOp(b_op.rd, -1) for b_op in b_ops)
    set_bound_ops = tuple(
        snitch.SsrSetDimensionBoundOp(new_b_op, dm, i)
        for (i, new_b_op) in zip(ints, new_b_ops, strict=True)
    )
    interleaved_b_set_bound_ops = tuple(
        x for t in zip(new_b_ops, set_bound_ops) for x in t
    )
    s_ops = tuple(riscv.LiOp(s.data) for s in reversed(strides.data))

    new_ops: list[Operation] = [
        *b_ops,
        *interleaved_b_set_bound_ops,
        *s_ops,
        snitch.SsrSetDimensionStrideOp(s_ops[0], dm, ints[0]),
        a_op := riscv.LiOp(0),
    ]

    for i in range(1, rank):
        a_inc_op = riscv.MulOp(new_b_ops[i - 1], s_ops[i - 1])
        new_a_op = riscv.AddOp(a_op, a_inc_op)
        stride_op = riscv.SubOp(s_ops[i], new_a_op)
        set_stride_op = snitch.SsrSetDimensionStrideOp(stride_op.rd, dm, ints[i])
        new_ops.extend((a_inc_op, new_a_op, stride_op, set_stride_op))
        a_op = new_a_op

    # Always reset the repetition count, even if it's the default
    new_ops.extend(
        (
            repeat_op := riscv.LiOp(repeat.data - 1),
            snitch.SsrSetStreamRepetitionOp(repeat_op.rd, dm),
        )
    )

    rewriter.insert_op(new_ops, InsertPoint.before(target_op))