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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
/*******************************************************************************
* Copyright 2018 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.
*******************************************************************************/

#include "math_utils.hpp"
#include "mkldnn_thread.hpp"
#include "simple_q10n.hpp"

#include "gemm/gemm.hpp"
#include "gemm_x8s8s32x_inner_product.hpp"

namespace mkldnn {
namespace impl {
namespace cpu {

using namespace math;
using namespace format_tag;
using namespace memory_tracking::names;

template<data_type_t src_type, data_type_t dst_type>
gemm_x8s8s32x_inner_product_fwd_t<src_type, dst_type>::pp_kernel_t::pp_kernel_t(
        const pd_t *pd, bool dst_is_acc)
    : ker_(nullptr), OC_(pd->OC())
    , bias_data_type_(data_type::undef), bias_data_type_size_(0)
    , scale_idx_mult_(0), do_bias_(false), do_relu_(false)
{
    using namespace types;

    scale_idx_mult_ = (pd->attr()->output_scales_.mask_ == (1 << 1));

    auto &post_ops = pd->attr()->post_ops_;
    do_relu_ = post_ops.len_ == 1;
    do_bias_ = pd->with_bias();
    bias_data_type_ = pd->desc()->bias_desc.data_type;
    if (do_bias_) {
        assert(bias_data_type_ != data_type::undef);
        bias_data_type_size_ = data_type_size(bias_data_type_);
    }

    if (!mayiuse(avx512_core))
        // use fallback code for older CPUs since they do not have optimized
        // x8s8s32 GEMM anyways. The configuration variables above are used by
        // the fallback code.
        return;
    else
        generate();
}

template<data_type_t src_type, data_type_t dst_type>
void gemm_x8s8s32x_inner_product_fwd_t<src_type, dst_type>::pp_kernel_t::generate()
{
    using namespace Xbyak;
    using namespace utils;

    // TODO: clean-up
    Reg64 reg_param = abi_param1;
    Reg64 reg_dst = rdx;
    Reg64 reg_acc = rax;
    Reg64 reg_bias = rbx;
    Reg64 reg_scales = rsi;

    Reg64 reg_len = r8;
    Reg64 reg_tmp = rcx; // intentional for shifting purposes
    Reg64 reg_oc_offset = r9;
    Reg64 reg_rem_mask = r10;
    Opmask kreg_rem_mask = k1;
    Opmask kreg_relu_cmp = k2;

    const size_t vlen = cpu_isa_traits<avx512_common>::vlen / sizeof(float);

    Zmm vreg_zero = Zmm(0);
    Zmm vreg_scale = Zmm(1);
    Zmm vreg_nslope = Zmm(2);

    auto vreg_dst = [&](int idx) { return Zmm(3 + idx * 2 + 0); };
    auto vreg_bias = [&](int idx) { return Zmm(3 + idx * 2 + 1); };

    preamble();

#define PARAM_OFF(x) offsetof(ker_args, x)
    mov(reg_dst, ptr[reg_param + PARAM_OFF(dst)]);
    mov(reg_acc, ptr[reg_param + PARAM_OFF(acc)]);
    mov(reg_bias, ptr[reg_param + PARAM_OFF(bias)]);
    mov(reg_scales, ptr[reg_param + PARAM_OFF(scales)]);
    mov(reg_len, ptr[reg_param + PARAM_OFF(len)]);
    mov(reg_oc_offset, ptr[reg_param + PARAM_OFF(oc_offset)]);
    vbroadcastss(vreg_nslope, ptr[reg_param + PARAM_OFF(nslope)]);
    if (scale_idx_mult_ == 0)
        vbroadcastss(vreg_scale, dword[reg_scales]);
#undef PARAM_OFF

    if (do_relu_ || dst_type == data_type::u8)
        vxorps(vreg_zero, vreg_zero, vreg_zero);

    // Load accumulated value, convert to float, apply bias (if any), scaling,
    // and relu (if any); then convert to destination type and store
    auto compute = [&](size_t offset, int idx, bool apply_mask) {
        auto acc_addr = ptr[reg_acc + offset * sizeof(acc_data_t)];

        if (scale_idx_mult_ > 0) {
            assert(scale_idx_mult_ == 1);
            auto scale_addr = ptr[reg_scales + offset * sizeof(float)];
            auto vreg_scale_ = vreg_scale;
            if (apply_mask)
                vreg_scale_ = vreg_scale_ | kreg_rem_mask;
            vmovups(vreg_scale, scale_addr);
        }

        auto vreg_dst_ = vreg_dst(idx);
        if (apply_mask)
            vreg_dst_ = vreg_dst_ | kreg_rem_mask;
        vcvtdq2ps(vreg_dst_, acc_addr);

        if (do_bias_) {
            auto bias_addr = ptr[reg_bias + offset * bias_data_type_size_];
            auto vreg_bias_ = vreg_bias(idx);
            if (apply_mask)
                vreg_bias_ = vreg_bias_ | kreg_rem_mask;

            switch (bias_data_type_) {
            case data_type::s8:
                vpmovsxbd(vreg_bias_, bias_addr);
                break;
            case data_type::u8:
                vpmovzxbd(vreg_bias_, bias_addr);
                break;
            case data_type::s32:
            case data_type::f32:
                vmovups(vreg_bias_, bias_addr);
                break;
            default: assert(!"unimplemented");
            }
            if (bias_data_type_ != data_type::f32)
                vcvtdq2ps(vreg_bias(idx), vreg_bias(idx));
            vaddps(vreg_dst(idx), vreg_dst(idx), vreg_bias(idx));
        }

        vmulps(vreg_dst(idx), vreg_dst(idx), vreg_scale);
        if (do_relu_) {
            vcmpps(kreg_relu_cmp, vreg_dst(idx), vreg_zero, _cmp_lt_os);
            vmulps(vreg_dst(idx) | kreg_relu_cmp, vreg_dst(idx), vreg_nslope);
        }

        if (dst_type == data_type::u8)
            vmaxps(vreg_dst(idx), vreg_dst(idx), vreg_zero);

        if (dst_type != data_type::f32) {
            vcvtps2dq(vreg_dst(idx), vreg_dst(idx));
        }

        auto dst_addr = ptr[reg_dst + offset * sizeof(dst_data_t)];
        switch (dst_type) {
        case data_type::s8:
            vpmovsdb(dst_addr, vreg_dst_);
            break;
        case data_type::u8:
            vpmovusdb(dst_addr, vreg_dst_);
            break;
        case data_type::f32:
        case data_type::s32:
            vmovups(dst_addr, vreg_dst_);
            break;
        default: assert(!"unimplemented");
        }
    };

    // Advance all pointers by an immediate
    auto advance_ptrs_imm = [&](size_t offset) {
        add(reg_dst, offset * sizeof(dst_data_t));
        add(reg_acc, offset * sizeof(acc_data_t));
        if (scale_idx_mult_) {
            assert(scale_idx_mult_ == 1);
            add(reg_scales, offset * sizeof(float));
        }
        if (do_bias_)
            add(reg_bias, offset * bias_data_type_size_);
    };

    // Advance all pointers by a value stored in a register
    auto advance_ptrs_reg = [&](Reg64 offset) {
        lea(reg_dst, ptr[reg_dst + offset * sizeof(dst_data_t)]);
        lea(reg_acc, ptr[reg_acc + offset * sizeof(acc_data_t)]);
        if (scale_idx_mult_) {
            assert(scale_idx_mult_ == 1);
            lea(reg_scales, ptr[reg_scales + offset * sizeof(float)]);
        }
        if (do_bias_)
            lea(reg_bias, ptr[reg_bias + offset * bias_data_type_size_]);
    };

    // Rewind pointers that point to data that is indixed by output channel
    // (bias or per-oc scaling factors)
    auto rewind_ptrs = [&]() {
        if (do_bias_)
            sub(reg_bias, OC_ * bias_data_type_size_);
        if (scale_idx_mult_) {
            assert(scale_idx_mult_ == 1);
            sub(reg_scales, OC_ * sizeof(float));
        }
    };

    //      <-------------------- OC ------------------------------->
    //
    // ^    +....................+----------------------------------+
    // |    :   not accessed     |          Prologue loop           |
    // |    +--------------------+----------------------------------+
    //      |                                                       |
    // M    |                 Main loop (unrolled)                  |
    // B    |                                                       |
    //      +--------------------------------+----------------------+
    // |    |       Epilogue loop            |      not accessed    :
    // v    +--------------------------------+......................+

    Label prologue_end;
    cmp(reg_oc_offset, 0);
    je(prologue_end, T_NEAR);

    // Prologue loop
    {
        mov(reg_tmp, OC_);
        sub(reg_tmp, reg_oc_offset);
        cmp(reg_tmp, reg_len);
        cmovg(reg_tmp, reg_len);
        sub(reg_len, reg_tmp);

        Label prologue_loop, prologue_loop_tail, prologue_loop_end;
        cmp(reg_tmp, vlen);
        jle(prologue_loop_tail, T_NEAR); // Skips for reg_tmp == 16 too (?)
        L(prologue_loop); {
            compute(0, 0, false);
            advance_ptrs_imm(vlen);
            sub(reg_tmp, vlen);
            cmp(reg_tmp, vlen);
            jge(prologue_loop, T_NEAR);
        }

        L(prologue_loop_tail);
        mov(reg_rem_mask, 1);
        shl(reg_rem_mask, cl); // cl == reg_tmp because reg_tmp <= vlen here
        sub(reg_rem_mask, 1);
        jz(prologue_loop_end, T_NEAR);

        kmovq(kreg_rem_mask, reg_rem_mask);
        compute(0, 0, true);
        advance_ptrs_reg(reg_tmp);

        L(prologue_loop_end);
        rewind_ptrs();
    }
    L(prologue_end);

    // Main loop
    Label main_loop_end;
    {
        cmp(reg_len, OC_);
        jle(main_loop_end, T_NEAR);

        Label main_loop;
        L(main_loop); {
            size_t def_unroll = 4;
            size_t max_unroll = 13;

            size_t OC_loop, OC_tail;
            if (OC_ < max_unroll * vlen) {
                // Fully unroll small loops
                OC_loop = 0;
                OC_tail = OC_;
            } else {
                OC_loop = vlen * def_unroll;
                OC_tail = OC_ % OC_loop;
            }

            assert(!!OC_loop || !!OC_tail);

            if (OC_tail % vlen) {
                int vlen_tail = OC_tail % vlen;
                unsigned tail_mask = (1 << vlen_tail) - 1;
                mov(reg_tmp, tail_mask);
                kmovq(kreg_rem_mask, reg_tmp);
            }

            if (OC_loop) {
                mov(reg_tmp, rnd_dn(OC_, OC_loop));
                Label oc_loop;
                L(oc_loop); {
                    for (size_t offset = 0; offset < OC_loop; offset += vlen)
                        compute(offset, offset / vlen, false);
                    advance_ptrs_imm(OC_loop);
                    sub(reg_tmp, OC_loop);
                    jnz(oc_loop);
                }
            }

            if (OC_tail) {
                for (size_t offset = 0; offset < OC_tail; offset += vlen) {
                    bool use_mask = (offset + vlen) > OC_tail;
                    compute(offset, offset / vlen, use_mask);
                }
                advance_ptrs_imm(OC_tail);
            }

            rewind_ptrs();
            sub(reg_len, OC_);
            cmp(reg_len, OC_);
            jge(main_loop, T_NEAR);
        }
    }
    L(main_loop_end);

    // Epilogue loop
    Label epilogue_end;
    {
        cmp(reg_len, 0);
        je(epilogue_end, T_NEAR);

        Label epilogue_loop, epilogue_loop_tail;
        cmp(reg_len, vlen);
        jle(epilogue_loop_tail, T_NEAR); // Skips for reg_len == 16 (?)
        L(epilogue_loop); {
            compute(0, 0, false);
            sub(reg_len, vlen);
            advance_ptrs_imm(vlen);
            cmp(reg_len, vlen);
            jge(epilogue_loop, T_NEAR);
        }

        L(epilogue_loop_tail);
        mov(reg_tmp, reg_len); // reg_tmp is rcx, and we need cl for the shift
        mov(reg_rem_mask, 1);
        shl(reg_rem_mask, cl); // reg_tmp == rcx and reg_tail < vlen == 16
        sub(reg_rem_mask, 1);
        jz(epilogue_end, T_NEAR);
        kmovq(kreg_rem_mask, reg_rem_mask);
        compute(0, 0, true);
    }

    L(epilogue_end);

    postamble();

    ker_ = getCode<decltype(ker_)>();
}

template<data_type_t src_type, data_type_t dst_type>
void gemm_x8s8s32x_inner_product_fwd_t<src_type, dst_type>::pp_kernel_t::operator ()(
        dst_data_t *dst, const acc_data_t *acc,
        const char *bias, const float *scales, float nslope,
        size_t start, size_t end)
{
    using math::get_bias;

    if (end <= start)
        return;

    if (ker_) {
        // JIT
        ker_args args;
        size_t oc_offset = start % OC_;
        args.dst = dst + start;
        args.acc = acc + start;
        args.bias = bias + oc_offset * bias_data_type_size_;
        args.scales = scales + scale_idx_mult_ * oc_offset;
        args.nslope = nslope;
        args.len = end - start;
        args.oc_offset = oc_offset;
        ker_(&args);
    } else {
        // Fallback
        size_t oc = start % OC_;
        for (size_t i = start; i < end; i++) {
            float d = (float)acc[i];
            float b = get_bias(bias, oc, bias_data_type_);
            d = d + b;
            d *= scales[oc * scale_idx_mult_];
            if (do_relu_ && d < 0)
                d *= nslope;
            dst[i] = qz_a1b0<float, dst_data_t>()(d);
            oc = (oc == OC_ - 1) ? 0 : oc + 1;
        }
    }
};

template <data_type_t src_type, data_type_t dst_type>
void gemm_x8s8s32x_inner_product_fwd_t<src_type, dst_type>::execute_forward(
        const exec_ctx_t &ctx) const {
    auto src = CTX_IN_MEM(const src_data_t *, MKLDNN_ARG_SRC);
    auto weights = CTX_IN_MEM(const wei_data_t *, MKLDNN_ARG_WEIGHTS);
    auto bias = CTX_IN_MEM(const char *, MKLDNN_ARG_BIAS);
    auto dst = CTX_OUT_MEM(dst_data_t *, MKLDNN_ARG_DST);

    const int MB = pd()->MB();
    const int OC = pd()->OC();

    bool wei_tr = memory_desc_matches_one_of_tag(
            *pd()->weights_md(), oiw, oihw, oidhw, oi);

    const int M = OC;
    const int N = MB;
    const int K = pd()->IC_total_padded();
    const int8_t off_a = 0, off_b = 0;
    const int32_t off_c = 0;

    const float *scales = pd()->attr()->output_scales_.scales_;

    const auto &post_ops = pd()->attr()->post_ops_;
    const bool do_relu = post_ops.len_ == 1;
    const float nslope = do_relu ? post_ops.entry_[0].eltwise.alpha : 0.f;

    acc_data_t *acc = pd()->dst_is_acc_
        ? (acc_data_t *)dst
        : scratchpad(ctx).template get<acc_data_t>(key_iprod_int_dat_in_acc_dt);

    const float onef = 1.0, zerof = 0.0;
    gemm_s8x8s32(wei_tr ? "T" : "N", "N", "F", &M, &N, &K, &onef, weights,
            wei_tr ? &K : &M, &off_a, src, &K, &off_b, &zerof, acc, &M, &off_c);

    if (!pd()->attr()->has_default_values() || !pd()->dst_is_acc_
            || pd()->with_bias()) {
        const bool force_sequential = MB * OC < 2000;
        parallel(force_sequential ? 1 : 0, [&](int ithr, int nthr) {
            size_t start, end;
            balance211((size_t)OC * MB, nthr, ithr, start, end);
            (*pp_kernel_)(dst, acc, bias, scales, nslope, start, end);
        });
    }
}

using namespace data_type;

template struct gemm_x8s8s32x_inner_product_fwd_t<u8, f32>;
template struct gemm_x8s8s32x_inner_product_fwd_t<u8, s32>;
template struct gemm_x8s8s32x_inner_product_fwd_t<u8, s8>;
template struct gemm_x8s8s32x_inner_product_fwd_t<u8, u8>;
template struct gemm_x8s8s32x_inner_product_fwd_t<s8, f32>;
template struct gemm_x8s8s32x_inner_product_fwd_t<s8, s32>;
template struct gemm_x8s8s32x_inner_product_fwd_t<s8, s8>;
template struct gemm_x8s8s32x_inner_product_fwd_t<s8, u8>;

}
}
}