• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Kernels in gemmlowp
2
3## Kernels provide an inner-loop implementation, and a format
4
5Here we assume familiarity with the concepts of kernels and of packing as
6explained in [design.md](design.md).
7
8gemmlowp is designed to be easily extensible to different architectures and
9other low-level details, while achieving high performance. Thus a line had to be
10drawn between the generic GEMM code and the specific parts that need to be
11manually designed for each architecture, etc. The design choice made in gemmlowp
12is to have easily swappable GEMM kernels.
13
14In itself, a GEMM kernel is just an implementation of the inner-most loop in a
15GEMM (That inner-most loop has to be over the 'depth' dimension so as to be able
16to accumulate into a small enough number of accumulators to fit in registers).
17
18Thus, by itself, a GEMM kernel should be just a function computing a block of
19GEMM.
20
21However, GEMM kernels may need to differ not just in how they implement this
22computation, but also in the format of data that they operate on. Indeed, in
23order to maximize the ratio of arithmetic instructions to memory access
24instructions, GEMM kernels want to handle blocks as wide as possible given the
25number of registers of the CPU architecture.
26
27Thus, in order to allow efficient specialization to diverse architectures,
28gemmlowp allows each GEMM kernel to dictate the format of data that it expects,
29in addition to providing its inner-loop implementation.
30
31The former is given by a 'Format' typedef, and the latter by a 'Run' method.
32
33A good example is to look at internal/kernel_neon.h, and specifically at the
34NEONKernel12x4Depth2 kernel, which specifies its format as
35
36```
37  typedef KernelFormat<KernelSideFormat<CellFormat<4, 2>, 3>,
38                       KernelSideFormat<CellFormat<4, 2>, 1> > Format;
39```
40
41The meaning of these terms is explained in the lengthy comment at the top of
42internal/kernel.h. Here, they mean that this kernel handles at each iteration
43(along the depth dimension): - 3 'cells' of size 4x2 each of the lhs, so a total
44lhs block of size 12x2 - 1 'cell' of size 2x4 of the rhs. In other words, this
45kernel handles 12 rows of the lhs and 4 columns of the rhs, and handles two
46levels of depth at once. The 'cells' and `CellFormat` detail the layout of these
4712x2 and 2x4 blocks.
48
49This kernel then loads these 12x2 and 2x4 blocks and computes the corresponding
5012x4 GEMM; for ease of reference let us paste the critical comment and code
51here:
52
53```
54"loop_NEONKernel12x4Depth2_%=:\n"
55
56// Overview of register layout:
57//
58// A 2x4 cell of Rhs is stored in 16bit in d0--d1 (q0).
59// A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in d2--d7
60// (q1--q3).
61// A 12x4 block of accumulators is stored in 32bit in q4--q15.
62//
63//                   +-----+-----+-----+-----+
64//                   |d0[0]|d0[1]|d0[2]|d0[3]|
65//              Rhs  +-----+-----+-----+-----+
66//                   |d1[0]|d1[1]|d1[2]|d1[3]|
67//                   +-----+-----+-----+-----+
68//
69//                   |     |     |     |     |
70//
71//    Lhs            |     |     |     |     |
72//
73//  +--+--+ - - - -  +-----+-----+-----+-----+
74//  |d2|d3|          | q4  | q5  | q6  | q7  |
75//  |d2|d3|          | q4  | q5  | q6  | q7  |
76//  |d2|d3|          | q4  | q5  | q6  | q7  |
77//  |d2|d3|          | q4  | q5  | q6  | q7  |
78//  +--+--+ - - - -  +-----+-----+-----+-----+
79//  |d4|d5|          | q8  | q9  | q10 | q11 |
80//  |d4|d5|          | q8  | q9  | q10 | q11 |
81//  |d4|d5|          | q8  | q9  | q10 | q11 |
82//  |d4|d5|          | q8  | q9  | q10 | q11 |
83//  +--+--+ - - - -  +-----+-----+-----+-----+
84//  |d6|d7|          | q12 | q13 | q14 | q15 |
85//  |d6|d7|          | q12 | q13 | q14 | q15 |
86//  |d6|d7|          | q12 | q13 | q14 | q15 |
87//  |d6|d7|          | q12 | q13 | q14 | q15 |
88//  +--+--+ - - - -  +-----+-----+-----+-----+
89//
90//                            Accumulator
91
92// Load 1 Rhs cell of size 2x4
93"vld1.8 {d0}, [%[rhs_ptr]:64]!\n"
94
95// Load 3 Lhs cells of size 4x2 each
96"vld1.8 {d2}, [%[lhs_ptr]:64]!\n"
97"vld1.8 {d4}, [%[lhs_ptr]:64]!\n"
98"vld1.8 {d6}, [%[lhs_ptr]:64]!\n"
99
100// Expand Lhs/Rhs cells to 16 bit.
101"vmovl.u8 q0, d0\n"
102"vmovl.u8 q1, d2\n"
103"vmovl.u8 q2, d4\n"
104"vmovl.u8 q3, d6\n"
105
106// Multiply-accumulate, level of depth 0
107"vmlal.u16 q4, d2, d0[0]\n"
108"vmlal.u16 q5, d2, d0[1]\n"
109"vmlal.u16 q6, d2, d0[2]\n"
110"vmlal.u16 q7, d2, d0[3]\n"
111"vmlal.u16 q8, d4, d0[0]\n"
112"vmlal.u16 q9, d4, d0[1]\n"
113"vmlal.u16 q10, d4, d0[2]\n"
114"vmlal.u16 q11, d4, d0[3]\n"
115"vmlal.u16 q12, d6, d0[0]\n"
116"vmlal.u16 q13, d6, d0[1]\n"
117"vmlal.u16 q14, d6, d0[2]\n"
118"vmlal.u16 q15, d6, d0[3]\n"
119
120// Multiply-accumulate, level of depth 1
121"vmlal.u16 q4, d3, d1[0]\n"
122"vmlal.u16 q5, d3, d1[1]\n"
123"vmlal.u16 q6, d3, d1[2]\n"
124"vmlal.u16 q7, d3, d1[3]\n"
125"vmlal.u16 q8, d5, d1[0]\n"
126"vmlal.u16 q9, d5, d1[1]\n"
127"vmlal.u16 q10, d5, d1[2]\n"
128"vmlal.u16 q11, d5, d1[3]\n"
129"vmlal.u16 q12, d7, d1[0]\n"
130"vmlal.u16 q13, d7, d1[1]\n"
131"vmlal.u16 q14, d7, d1[2]\n"
132"vmlal.u16 q15, d7, d1[3]\n"
133
134// Loop. Decrement loop index (depth) by 2, since we just handled 2
135// levels of depth (Kernel::kDepth=2).
136"subs %[run_depth], #2\n"
137"bne loop_NEONKernel12x4Depth2_%=\n"
138```
139
140## Packing code adapts to the format chosen by the kernel
141
142As explained in [design.md](design.md), gemmlowp starts by packing blocks of the
143lhs and rhs matrices for optimally efficient traversal by the kernel. This
144depends on fine details of the kernel format, in ways that can only be
145efficiently handled by knowing these kernel format details at compile-time.
146
147This is the reason why all the code in [internal/pack.h](../internal/pack.h) is
148templated in the corresponding kernel format.
149
150The code in internal/pack.h isn't tightly optimized by itself, but it is
151structured in such a way that the critical code is in a template,
152`PackingRegisterBlock`, that can easily be specialized to override the slow
153generic code with fast specific packing code for specific formats, on specific
154platforms.
155
156See [internal/pack_neon.h](../internal/pack_neon.h) which provides NEON
157specializations of the packing code for the particular kernel formats that are
158used by the NEON kernels in [internal/kernel_neon.h](../internal/kernel_neon.h).
159
160## Wrapping up: how to optimize gemmlowp for a CPU architecture
161
162In conclusion, the key feature of gemmlowp when it comes to efficiently
163supporting a specific CPU architecture, is that it allows to freely replace the
164inner loop of the GEMM by providing one's own GEMM kernel, which is also free to
165dictate its required data layout; each data layout then also needs optimized
166packing code. The steps are thus:
167
1681.  Freely design a GEMM kernel with a freely chosen data layout.
1692.  Implement the GEMM kernel, similar to
170    [internal/kernel_neon.h](../internal/kernel_neon.h).
1713.  Implement the optimized packing code, similar to
172    [internal/pack_neon.h](../internal/pack_neon.h).
173