edcb7162ac68f3014acda5b1bf6dcac7d4b41267
1 # SPDX-License-Identifier: LGPL-2.1-or-later
2 # See Notices.txt for copyright information
5 Copyright (C) 2020 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
6 Copyright (C) 2020 Michael Nolan <mtnolan2640@gmail.com>
8 dynamically partitionable shifter. Unlike part_shift_scalar, both
9 operands can be partitioned
13 * http://libre-riscv.org/3d_gpu/architecture/dynamic_simd/shift/
14 * http://bugs.libre-riscv.org/show_bug.cgi?id=173
16 from nmigen
import Signal
, Module
, Elaboratable
, Cat
, Mux
, C
17 from ieee754
.part_mul_add
.partpoints
import PartitionPoints
21 class PartitionedDynamicShift(Elaboratable
):
22 def __init__(self
, width
, partition_points
):
24 self
.partition_points
= PartitionPoints(partition_points
)
26 self
.a
= Signal(width
)
27 self
.b
= Signal(width
)
28 self
.output
= Signal(width
)
30 def elaborate(self
, platform
):
34 gates
= Signal(self
.partition_points
.get_max_partition_count(width
)-1)
35 comb
+= gates
.eq(self
.partition_points
.as_sig())
38 keys
= list(self
.partition_points
.keys()) + [self
.width
]
42 # break out both the input and output into partition-stratified blocks
49 for i
in range(len(keys
)):
51 widths
.append(width
- start
)
52 a_intervals
.append(self
.a
[start
:end
])
53 b_intervals
.append(self
.b
[start
:end
])
54 out_intervals
.append(self
.output
[start
:end
])
55 intervals
.append([start
,end
])
58 # Instead of generating the matrix described in the wiki, I
59 # instead calculate the shift amounts for each partition, then
60 # calculate the partial results of each partition << shift
61 # amount. On the wiki, the following table is given for output #3:
63 # 0 0 0 | a0b0[31:24] | a1b0[23:16] | a2b0[15:8] | a3b0[7:0]
64 # 0 0 1 | a0b0[31:24] | a1b1[23:16] | a2b1[15:8] | a3b1[7:0]
65 # 0 1 0 | a0b0[31:24] | a1b0[23:16] | a2b2[15:8] | a3b2[7:0]
66 # 0 1 1 | a0b0[31:24] | a1b1[23:16] | a2b2[15:8] | a3b2[7:0]
67 # 1 0 0 | a0b0[31:24] | a1b0[23:16] | a2b0[15:8] | a3b3[7:0]
68 # 1 0 1 | a0b0[31:24] | a1b1[23:16] | a2b1[15:8] | a3b3[7:0]
69 # 1 1 0 | a0b0[31:24] | a1b0[23:16] | a2b2[15:8] | a3b3[7:0]
70 # 1 1 1 | a0b0[31:24] | a1b1[23:16] | a2b2[15:8] | a3b3[7:0]
72 # Each output for o3 is given by a3bx and the partial results
73 # for o2 (namely, a2bx, a1bx, and a0b0). If I calculate the
74 # partial results [a0b0, a1bx, a2bx, a3bx], I can use just
75 # those partial results to calculate a0, a1, a2, and a3
77 partial_results
.append(a_intervals
[0] << b_intervals
[0])
78 element
= b_intervals
[0]
79 for i
in range(1, len(out_intervals
)):
81 element
= Mux(gates
[i
-1], b_intervals
[i
], element
)
83 # This calculates which partition of b to select the
84 # shifter from. According to the table above, the
85 # partition to select is given by the highest set bit in
86 # the partition mask, this calculates that with a mux
90 # This computes the partial results table
91 shifter
= Signal(8, name
="shifter%d" % i
)
92 comb
+= shifter
.eq(element
)
93 partial
= Signal(width
, name
="partial%d" % i
)
94 comb
+= partial
.eq(a_intervals
[i
] << shifter
)
96 partial_results
.append(partial
)
100 # This calculates the outputs o0-o3 from the partial results
103 result
= partial_results
[0]
104 out
.append(result
[s
:e
])
105 for i
in range(1, len(out_intervals
)):
106 start
, end
= (intervals
[i
][0], width
)
107 result
= partial_results
[i
] | \
108 Mux(gates
[i
-1], 0, result
[intervals
[0][1]:])[:end
-start
]
109 print("select: [%d:%d]" % (start
, end
))
110 res
= Signal(width
, name
="res%d" % i
)
111 comb
+= res
.eq(result
)
115 comb
+= self
.output
.eq(Cat(*out
))