Unpacker.cu 6.31 KB
Newer Older
Ewan Barr's avatar
Ewan Barr committed
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
#include "psrdada_cpp/effelsberg/edd/Unpacker.cuh"
#include "psrdada_cpp/cuda_utils.hpp"

#define EDD_NTHREADS_UNPACK 512

namespace psrdada_cpp {
namespace effelsberg {
namespace edd {
namespace kernels {

__device__ __forceinline__ uint64_t swap64(uint64_t x)
{
    uint64_t result;
    uint2 t;
    asm("mov.b64 {%0,%1},%2; \n\t"
        : "=r"(t.x), "=r"(t.y) : "l"(x));
    t.x = __byte_perm(t.x, 0, 0x0123);
    t.y = __byte_perm(t.y, 0, 0x0123);
    asm("mov.b64 %0,{%1,%2}; \n\t"
        : "=l"(result) : "r"(t.y), "r"(t.x));
    return result;
}

__global__
void unpack_edd_12bit_to_float32(uint64_t* __restrict__ in, float* __restrict__ out, int n)
{
    /**
     * Note: This kernels will not work with more than 512 threads.
     */
    __shared__ volatile float tmp_out[NTHREADS_UNPACK * 16];
    __shared__ volatile uint64_t tmp_in[NTHREADS_UNPACK * 3];
    int block_idx = blockIdx.x;
    uint64_t val;
    uint64_t rest;
    volatile float* sout = tmp_out + (16 * threadIdx.x);
    for (int idx = blockIdx.x * blockDim.x + threadIdx.x;
        (3 * idx + 2) < n;
        idx+=gridDim.x*blockDim.x)
    {
        //Read to shared memeory
        int block_read_start = block_idx * NTHREADS_UNPACK * 3;
        tmp_in[threadIdx.x]                = in[block_read_start + threadIdx.x];
        tmp_in[NTHREADS_UNPACK + threadIdx.x]     = in[block_read_start + NTHREADS_UNPACK + threadIdx.x];
        tmp_in[NTHREADS_UNPACK * 2 + threadIdx.x] = in[block_read_start + NTHREADS_UNPACK * 2 + threadIdx.x];
        __syncthreads();
        val  = swap64(tmp_in[3*threadIdx.x]);
        sout[0] = (float)((int64_t)(( 0xFFF0000000000000 & val) <<  0) >> 52);
        sout[1] = (float)((int64_t)(( 0x000FFF0000000000 & val) << 12) >> 52);
        sout[2] = (float)((int64_t)(( 0x000000FFF0000000 & val) << 24) >> 52);
        sout[3] = (float)((int64_t)(( 0x000000000FFF0000 & val) << 36) >> 52);
        sout[4] = (float)((int64_t)(( 0x000000000000FFF0 & val) << 48) >> 52);
        rest    =                   ( 0x000000000000000F & val) << 60;
        val  = swap64(tmp_in[3*threadIdx.x + 1]);
        sout[5] = (float)((int64_t)((( 0xFF00000000000000 & val) >> 4) | rest) >> 52);
        sout[6] = (float)((int64_t)((  0x00FFF00000000000 & val) << 8)  >> 52);
        sout[7] = (float)((int64_t)((  0x00000FFF00000000 & val) << 20) >> 52);
        sout[8] = (float)((int64_t)((  0x00000000FFF00000 & val) << 32) >> 52);
        sout[9] = (float)((int64_t)((  0x00000000000FFF00 & val) << 44) >> 52);
        rest    =                   (  0x00000000000000FF & val) << 56;
        val  = swap64(tmp_in[3*threadIdx.x + 2]);
        sout[10] = (float)((int64_t)((( 0xF000000000000000 & val) >>  8) | rest) >> 52);
        sout[11] = (float)((int64_t)((  0x0FFF000000000000 & val) <<  4) >> 52);
        sout[12] = (float)((int64_t)((  0x0000FFF000000000 & val) << 16) >> 52);
        sout[13] = (float)((int64_t)((  0x0000000FFF000000 & val) << 28) >> 52);
        sout[14] = (float)((int64_t)((  0x0000000000FFF000 & val) << 40) >> 52);
        __syncthreads();
        int block_write_start = block_idx * NTHREADS_UNPACK * 16;
        for (int ii = threadIdx.x; ii < 16 * NTHREADS_UNPACK; ii += blockDim.x)
        {
            out[block_write_start + ii] = tmp_out[ii];
        }
        block_idx += gridDim.x;
    }
}

__global__
void unpack_edd_8bit_to_float32(uint64_t* __restrict__ in, float* __restrict__ out, int n);
{
    /**
     * Note: This kernels will not work with more than 512 threads.
     */
    __shared__ volatile float tmp_out[NTHREADS_UNPACK * 8];
    int block_idx = blockIdx.x;
    uint64_t val;
    volatile float* sout = tmp_out + (8 * threadIdx.x);

    for (int idx = blockIdx.x * blockDim.x + threadIdx.x ; idx < n ; idx+=gridDim.x*blockDim.x)
    {
        int block_read_start = block_idx * NTHREADS_UNPACK;
        val = swap64(in[block_read_start + threadIdx.x]);
        sout[0] = (float)((int64_t)(( 0xFF00000000000000 & val) <<  0) >> 56);
        sout[1] = (float)((int64_t)(( 0x00FF000000000000 & val) <<  8) >> 56);
        sout[2] = (float)((int64_t)(( 0x0000FF0000000000 & val) << 16) >> 56);
        sout[3] = (float)((int64_t)(( 0x000000FF00000000 & val) << 24) >> 56);
        sout[4] = (float)((int64_t)(( 0x00000000FF000000 & val) << 32) >> 56);
        sout[5] = (float)((int64_t)(( 0x0000000000FF0000 & val) << 40) >> 56);
        sout[6] = (float)((int64_t)(( 0x000000000000FF00 & val) << 48) >> 56);
        sout[7] = (float)((int64_t)(( 0x00000000000000FF & val) << 56) >> 56);
        __syncthreads();
        int block_write_start = block_idx * NTHREADS_UNPACK * 8;
        for (int ii = threadIdx.x; ii < 8 * NTHREADS_UNPACK; ii+=blockDim.x)
        {
            out[block_write_start+ii] = tmp_out[ii];
        }
        block_idx += gridDim.x;
    }
}

} //namespace kernels


Unpacker::Unpacker(cudaStream_t stream)
    : _stream(stream)
{

}

Unpacker::~Unpacker()
{

}

template <>
void Unpacker::unpack<12>(InputType const& input, OutputType& output)
{
    BOOST_LOG_TRIVIAL(debug) << "Unpacking 12-bit data";
    std::size_t output_size = input.size() * 16 / 3;
    BOOST_LOG_TRIVIAL(debug) << "Resizing output buffer to " << output_size << " elements";
    output.resize(output_size);
    int nblocks = input.size() / EDD_NTHREADS_UNPACK;
    InputType::value_type const* input_ptr = thrust::raw_pointer_cast(input.data())
    OutputType::value_type* output_ptr = thrust::raw_pointer_cast(output.data())
    kernels::unpack_edd_12bit_to_float32<<< nblocks, EDD_NTHREADS_UNPACK, 0, _stream>>>(
            input_ptr, output_ptr, input.size());
    CUDA_ERROR_CHECK(cudaStreamSynchronize(_stream));
}

template <>
void Unpacker::unpack<8>(InputType const& input, OutputType& output)
{
    BOOST_LOG_TRIVIAL(debug) << "Unpacking 12-bit data";
    std::size_t output_size = input.size() * 8;
    BOOST_LOG_TRIVIAL(debug) << "Resizing output buffer to " << output_size << " elements";
    output.resize(output_size);
    int nblocks = input.size() / EDD_NTHREADS_UNPACK;
    InputType::value_type const* input_ptr = thrust::raw_pointer_cast(input.data())
    OutputType::value_type* output_ptr = thrust::raw_pointer_cast(output.data())
    kernels::unpack_edd_8bit_to_float32<<< nblocks, EDD_NTHREADS_UNPACK, 0, _stream>>>(
            input_ptr, output_ptr, input.size());
    CUDA_ERROR_CHECK(cudaStreamSynchronize(_stream));
}

} //namespace edd
} //namespace effelsberg
} //namespace psrdada_cpp