onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation 
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

/*******************************************************************************
* Copyright 2024 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#ifndef GPU_INTEL_REDUCTION_JIT_GENERATOR_HPP
#define GPU_INTEL_REDUCTION_JIT_GENERATOR_HPP

#include "common/c_types_map.hpp"
#include "common/nstl.hpp"
#include "gpu/intel/compute/device_info.hpp"
#include "gpu/intel/jit/emulated_generator.hpp"
#include "gpu/intel/jit/generator.hpp"
#include "gpu/intel/jit/reduction_injector.hpp"
#include "ngen_core.hpp"
#include "ngen_interface.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace reduction {
namespace jit {

using namespace intel::jit;

template <gpu_gen_t hw>
class generator_t : public emulated_generator_t<hw> {
protected:
    NGEN_FORWARD_ELF(hw)
    FORWARD_EMULATION(hw);

public:
    generator_t(const compute::device_info_t &device_info, impl::alg_kind_t alg,
            dim_t stride, dim_t iters, int nregs)
        : emulated_generator_t<hw>(
                  device_info, {GENERATOR_NAME, GENERATOR_LINE}) {
        constexpr auto GlobalPtr = ngen::ExternalArgumentType::GlobalPtr;

        // Number of dst elements computed per thread
        const int grf_bytes = ngen::GRF::bytes(hw);
        const int dt_size = sizeof(float);
        int simd = grf_bytes / dt_size;

        newArgument("src_ptr", GlobalPtr);
        newArgument("dst_ptr", GlobalPtr);
        setDefaultAutoSWSB();
        requireSIMD(simd);

        requireLocalSize();
        externalName("ngen_jit_reduction");
        finalizeInterface();

        prologue();
        setDefaultNoMask();

        ra().claim(r0);
        ngen::Subregister tg_idx0 = r0.ud(1);
        ngen::Subregister tg_idx1 = r0.ud(6);
        ngen::Subregister tid = r0.ud(2).b(0);
        ngen::Subregister tg_size0 = getLocalSize(0).uw();
        ngen::Subregister src_ptr = getArgument("src_ptr").uq();
        ngen::Subregister dst_ptr = getArgument("dst_ptr").uq();
        ra().claim(src_ptr);
        ra().claim(dst_ptr);
        ra().claim(tg_size0);

        // SRC offset: tid*grf_bytes*nregs + tg_idx0*tg_size0*dt_size*nregs + tg_idx1*stride*iters*dt_size
        // DST offset: tid*grf_bytes*nregs + tg_idx0*tg_size0*dt_size*nregs + tg_idx1*stride*dt_size
        ngen::Subregister inner_off = ra().alloc_sub(ngen::DataType::ud);
        ngen::Subregister outer_off = ra().alloc_sub(ngen::DataType::ud);
        emul(1, inner_off, tg_idx0, tg_size0);
        emul(1, inner_off, inner_off, dt_size * nregs);
        emad(1, inner_off, inner_off, tid, nregs * grf_bytes);

        emul(1, outer_off, tg_idx1, stride * dt_size);

        ngen::GRF src_addr = ra().alloc().uq();
        ngen::GRF dst_addr = ra().alloc().uq();
        emad(1, src_addr, inner_off, outer_off, iters);
        eadd(1, src_addr, src_addr, src_ptr);
        eadd3(1, dst_addr, dst_ptr, inner_off, outer_off);
        ra().release(inner_off);
        ra().release(outer_off);

        ngen::GRFRange acc = ra().alloc_range(nregs);
        reduction_injector_f32_t<typename generator_t<hw>::RootCodeGenerator>
                reduce(*this, alg, ra(), device_info.stepping_id());
        reduce.compute(src_addr, acc, stride, iters);
        ra().release(src_addr);

        finalize(simd, alg, acc, iters);

        estore(dst_addr, acc);

        ra().release(acc);
        ra().release(dst_addr);

        ra().release(r0);
        ra().release(src_ptr);
        ra().release(dst_ptr);
        ra().release(tg_size0);
        epilogue();
    }

protected:
    // Store data from a contiguous range of registers into a contiguous
    // range in global memory (block store)
    void estore(const ngen::GRF &base_dst_addr, const ngen::GRFRange &src) {
        const int grf_bytes = ngen::GRF::bytes(hw);
        int nregs = src.getLen();
        bool force_legacy = gpu_utils::dev_getenv(
                "jit_reduction_force_legacy_send", false);
        bool use_legacy = force_legacy || hw < ngen::HW::XeHPG;
        const int max_store_size = use_legacy ? 128 : 512;
        gpu_assert(max_store_size % grf_bytes == 0) << "Unexpected store size";
        const int max_store_regs = max_store_size / grf_bytes;

        // Load in chunks
        int reg_start = 0;
        while (reg_start < nregs) {
            int store_regs = nstl::min(max_store_regs, nregs - reg_start);
            // Compute the src address
            ngen::GRF addr = ra().alloc().uq();
            eadd(1, addr, base_dst_addr, reg_start * grf_bytes);
            if (use_legacy) {
                // Reduce store_regs according to valid store sizes
                const int oword_per_grf = grf_bytes / 16;
                for (auto store_owords : {8, 4, 2, 1}) {
                    if (store_owords / oword_per_grf > store_regs) continue;
                    store_regs = store_owords / oword_per_grf;
                    break;
                }

                // Do the store
                auto dt = ngen::aligned_block_oword(store_regs * oword_per_grf);
                store(1, dt, A64, addr, src[reg_start]);
            } else {
                // Reduce store_regs according to valid store sizes
                const int d64_per_grf = grf_bytes / 8;
                for (auto store_d64s : {64, 32, 16, 8, 4, 3, 2, 1}) {
                    if (store_d64s / d64_per_grf > store_regs) continue;
                    store_regs = store_d64s / d64_per_grf;
                    break;
                }

                // Do the store
                ngen::DataSpecLSC lscspec = ngen::CacheSettingsLSC::L1UC_L3WB;
                lscspec |= ngen::block(
                        ngen::DataSizeLSC::D64, store_regs * d64_per_grf);
                store.ugm(1, lscspec, A64, addr, src[reg_start]);
            }
            reg_start += store_regs;
            ra().release(addr);
        }
    }

    void finalize(int simd, impl::alg_kind_t alg, const ngen::GRFRange &acc,
            dim_t iters) {
        int nregs = acc.getLen();
        for (int i = 0; i < nregs; i++) {
            switch (alg) {
                case alg_kind::reduction_mean:
                    mul(simd, acc[i].f(), acc[i].f(), 1.0f / iters);
                    break;
                default: break;
            }
        }
    }
};

} // namespace jit
} // namespace reduction
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif // GPU_INTEL_REDUCTION_JIT_GENERATOR_HPP