blockwise_gemm.hip.hpp 44.2 KB
Newer Older
Chao Liu's avatar
Chao Liu committed
1
#pragma once
2
#include "threadwise_gemm.hip.hpp"
Chao Liu's avatar
Chao Liu committed
3

Jing Zhang's avatar
inline  
Jing Zhang committed
4
5
extern "C" __attribute__((address_space(3))) void* __to_local(void* p) [[hc]];

Chao Liu's avatar
Chao Liu committed
6
template <index_t BlockSize,
Chao Liu's avatar
Chao Liu committed
7
8
9
10
11
12
          class BlockMatrixA,
          class BlockMatrixB,
          class ThreadMatrixC,
          bool TransA,
          bool TransB,
          bool TransC,
Chao Liu's avatar
Chao Liu committed
13
14
15
          index_t KPerThreadLoop,
          index_t MThreadPerCluster,
          index_t NThreadPerCluster,
Chao Liu's avatar
Chao Liu committed
16
          bool DistributeThreadAlongColumnFirst>
17
struct BlockwiseGemmBlockABlockBThreadC
Chao Liu's avatar
Chao Liu committed
18
{
Chao Liu's avatar
Chao Liu committed
19
20
    index_t mMyThreadOffsetA = 0;
    index_t mMyThreadOffsetB = 0;
Chao Liu's avatar
Chao Liu committed
21
22
23

    struct MatrixIndex
    {
Chao Liu's avatar
Chao Liu committed
24
25
        index_t row;
        index_t col;
Chao Liu's avatar
Chao Liu committed
26
27
    };

28
    __device__ BlockwiseGemmBlockABlockBThreadC()
Chao Liu's avatar
Chao Liu committed
29
    {
Chao Liu's avatar
Chao Liu committed
30
31
        constexpr auto a_block_mtx = BlockMatrixA{};
        constexpr auto b_block_mtx = BlockMatrixB{};
Chao Liu's avatar
Chao Liu committed
32

Chao Liu's avatar
Chao Liu committed
33
        const auto c_thread_mtx_index = GetBeginOfThreadMatrixC(get_thread_local_1d_id());
Chao Liu's avatar
Chao Liu committed
34

Chao Liu's avatar
Chao Liu committed
35
36
        mMyThreadOffsetA = (!TransA) ? a_block_mtx.Get1dIndex(c_thread_mtx_index.row, 0)
                                     : a_block_mtx.Get1dIndex(0, c_thread_mtx_index.row);
Chao Liu's avatar
Chao Liu committed
37

Chao Liu's avatar
Chao Liu committed
38
39
        mMyThreadOffsetB = (!TransB) ? b_block_mtx.Get1dIndex(0, c_thread_mtx_index.col)
                                     : b_block_mtx.Get1dIndex(c_thread_mtx_index.col, 0);
Chao Liu's avatar
Chao Liu committed
40
41
42
43
44
45
46
47
48
49
50

#if 0
        if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
        {
            print_ConstantMatrixDescriptor(BlockMatrixA{}, "a_block_mtx: ");
            print_ConstantMatrixDescriptor(BlockMatrixB{}, "b_block_mtx: ");
            print_ConstantMatrixDescriptor(ThreadMatrixC{}, "c_thread_mtx: ");

            printf("%u %u, %u %u %u, %u %u\n",
                   get_block_1d_id(),
                   get_thread_local_1d_id(),
Chao Liu's avatar
Chao Liu committed
51
52
53
                   c_thread_mtx_index.batch,
                   c_thread_mtx_index.row,
                   c_thread_mtx_index.col,
Chao Liu's avatar
Chao Liu committed
54
55
56
57
58
59
                   mMyThreadOffsetA,
                   mMyThreadOffsetB);
        }
#endif
    }

Chao Liu's avatar
Chao Liu committed
60
    __device__ MatrixIndex GetBeginOfThreadMatrixC(index_t thread_id) const
Chao Liu's avatar
Chao Liu committed
61
62
63
64
    {

        if(TransA && (!TransB) && (!TransC))
        {
Chao Liu's avatar
Chao Liu committed
65
66
            constexpr auto a_block_mtx = BlockMatrixA{};
            constexpr auto b_block_mtx = BlockMatrixB{};
Chao Liu's avatar
Chao Liu committed
67
68
69
70

            static_assert(a_block_mtx.NRow() == b_block_mtx.NRow(),
                          "wrong! k dimension not consistent!");

Chao Liu's avatar
Chao Liu committed
71
72
            constexpr index_t MPerBlock = a_block_mtx.NCol();
            constexpr index_t NPerBlock = b_block_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
73

Chao Liu's avatar
Chao Liu committed
74
            constexpr auto c_thread_mtx = ThreadMatrixC{};
Chao Liu's avatar
Chao Liu committed
75
76

            // divide thread work
Chao Liu's avatar
Chao Liu committed
77
78
            constexpr index_t MPerThread = c_thread_mtx.NRow();
            constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
79
80
81
82
83
84
85

            static_assert(MPerBlock % (MPerThread * MThreadPerCluster) == 0,
                          "MPerBlock % (MPerThread * MThreadPerCluster) != 0");

            static_assert(NPerBlock % (NPerThread * NThreadPerCluster) == 0,
                          "NPerBlock % (NPerThread * NThreadPerCluster) != 0");

Chao Liu's avatar
Chao Liu committed
86
            constexpr index_t MClusterWork =
Chao Liu's avatar
Chao Liu committed
87
88
                (MPerBlock + MPerThread * MThreadPerCluster - 1) / (MPerThread * MThreadPerCluster);

Chao Liu's avatar
Chao Liu committed
89
            constexpr index_t NClusterWork =
Chao Liu's avatar
Chao Liu committed
90
91
                (NPerBlock + NPerThread * NThreadPerCluster - 1) / (NPerThread * NThreadPerCluster);

Chao Liu's avatar
Chao Liu committed
92
93
94
            static_assert(BlockSize ==
                              (MClusterWork * MThreadPerCluster) *
                                  (NClusterWork * NThreadPerCluster),
Chao Liu's avatar
Chao Liu committed
95
96
97
98
                          "wrong! wrong BlockSize");

            if(DistributeThreadAlongColumnFirst)
            {
Chao Liu's avatar
Chao Liu committed
99
                const index_t cluster_work_block_id =
Chao Liu's avatar
Chao Liu committed
100
101
                    thread_id / (MThreadPerCluster * NThreadPerCluster);

Chao Liu's avatar
Chao Liu committed
102
                const index_t thread_work_cluster_id =
Chao Liu's avatar
Chao Liu committed
103
104
                    thread_id - cluster_work_block_id * (MThreadPerCluster * NThreadPerCluster);

Chao Liu's avatar
Chao Liu committed
105
106
                const index_t m_cluster_work_block_id = cluster_work_block_id / NClusterWork;
                const index_t n_cluster_work_block_id =
Chao Liu's avatar
tune  
Chao Liu committed
107
                    cluster_work_block_id - m_cluster_work_block_id * NClusterWork;
Chao Liu's avatar
Chao Liu committed
108

Chao Liu's avatar
Chao Liu committed
109
110
                const index_t m_thread_work_cluster_id = thread_work_cluster_id / NThreadPerCluster;
                const index_t n_thread_work_cluster_id =
Chao Liu's avatar
Chao Liu committed
111
112
113
114
115
116
                    thread_work_cluster_id - m_thread_work_cluster_id * NThreadPerCluster;

#if 0
                if(get_block_1d_id() == 0)
                {
                    printf("%u %u, \t"
Chao Liu's avatar
tune  
Chao Liu committed
117
                           "MClusterWork %u MThreadPerCluster %u NClusterWork %u NThreadPerCluster %u \t"
Chao Liu's avatar
Chao Liu committed
118
119
120
121
                           "m_cluster_work_block_id %u n_cluster_work_block_id %u \t"
                           "m_thread_work_cluster_id %u n_thread_work_cluster_id %u \t"
                            "\n",
                            get_block_1d_id(), get_thread_local_1d_id(),
Chao Liu's avatar
tune  
Chao Liu committed
122
                            MClusterWork, MThreadPerCluster, NClusterWork, NThreadPerCluster,
Chao Liu's avatar
Chao Liu committed
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
                            m_cluster_work_block_id, n_cluster_work_block_id,
                            m_thread_work_cluster_id, n_thread_work_cluster_id);
                }
#endif

                return MatrixIndex{m_cluster_work_block_id * (MThreadPerCluster * MPerThread) +
                                       m_thread_work_cluster_id * MPerThread,
                                   n_cluster_work_block_id * (NThreadPerCluster * NPerThread) +
                                       n_thread_work_cluster_id * NPerThread};
            }
            else
            {
                // not implemented
                assert(false);
            }
        }
        else
        {
            // not implemented
            assert(false);
        }
    }

Chao Liu's avatar
Chao Liu committed
146
    // this should be optimized away if input is known
Chao Liu's avatar
Chao Liu committed
147
148
    __device__ static MatrixIndex GetDistanceFromBeginOfThreadMatrixC(index_t m_in_c,
                                                                      index_t n_in_c)
Chao Liu's avatar
Chao Liu committed
149
150
151
152
    {
        return MatrixIndex{m_in_c, n_in_c};
    }

Chao Liu's avatar
Chao Liu committed
153
    template <class FloatA, class FloatB, class FloatC, class Accumulator>
Chao Liu's avatar
Chao Liu committed
154
155
156
    __device__ void Run(const FloatA* __restrict__ p_a_block,
                        const FloatB* __restrict__ p_b_block,
                        FloatC* __restrict__ p_c_thread,
Chao Liu's avatar
Chao Liu committed
157
158
159
160
                        Accumulator f_accum) const
    {
        if(TransA && (!TransB) && (!TransC))
        {
Chao Liu's avatar
Chao Liu committed
161
162
            constexpr auto True  = integral_constant<bool, true>{};
            constexpr auto False = integral_constant<bool, false>{};
Chao Liu's avatar
Chao Liu committed
163

Chao Liu's avatar
Chao Liu committed
164
165
166
            constexpr auto a_block_mtx  = BlockMatrixA{};
            constexpr auto b_block_mtx  = BlockMatrixB{};
            constexpr auto c_thread_mtx = ThreadMatrixC{};
Chao Liu's avatar
Chao Liu committed
167

Chao Liu's avatar
Chao Liu committed
168
            constexpr index_t KPerBlock = a_block_mtx.NRow(); // A is transposed
Chao Liu's avatar
Chao Liu committed
169

Chao Liu's avatar
Chao Liu committed
170
171
            constexpr index_t MPerThread = c_thread_mtx.NRow();
            constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
172
173

            // a is transposed, b is not
Chao Liu's avatar
Chao Liu committed
174
175
            constexpr auto a_thread_mtx =
                make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
Chao Liu's avatar
Chao Liu committed
176

Chao Liu's avatar
Chao Liu committed
177
178
            constexpr auto b_thread_mtx =
                make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
Chao Liu's avatar
Chao Liu committed
179
180
181
182
183

            FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
            FloatB p_b_thread[b_thread_mtx.GetElementSpace()];

            // loop over k
Chao Liu's avatar
Chao Liu committed
184
            for(index_t k_begin = 0; k_begin < KPerBlock; k_begin += KPerThreadLoop)
Chao Liu's avatar
Chao Liu committed
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
            {
                threadwise_matrix_copy(a_block_mtx,
                                       p_a_block + mMyThreadOffsetA +
                                           k_begin * a_block_mtx.RowStride(),
                                       a_thread_mtx,
                                       p_a_thread,
                                       a_thread_mtx.GetLengths());

                threadwise_matrix_copy(b_block_mtx,
                                       p_b_block + mMyThreadOffsetB +
                                           k_begin * b_block_mtx.RowStride(),
                                       b_thread_mtx,
                                       p_b_thread,
                                       b_thread_mtx.GetLengths());

                threadwise_gemm(a_thread_mtx,
                                True,
                                p_a_thread,
                                b_thread_mtx,
                                False,
                                p_b_thread,
                                c_thread_mtx,
                                False,
                                p_c_thread,
                                f_accum);
            }
        }
    }
};
Chao Liu's avatar
Chao Liu committed
214
215
216

// if following number are power of 2, index calculation shall be greatly reduced:
//    MPerThreadSubC, NPerThreadSubC, MLevel0Cluster, NLevel0Cluster, MLevel1Cluster, NLevel1Cluster
Chao Liu's avatar
Chao Liu committed
217
template <index_t BlockSize,
Chao Liu's avatar
Chao Liu committed
218
219
220
          class BlockMatrixA,
          class BlockMatrixB,
          class ThreadMatrixC,
Chao Liu's avatar
Chao Liu committed
221
222
223
224
225
226
227
          index_t MPerThreadSubC,
          index_t NPerThreadSubC,
          index_t MLevel0Cluster,
          index_t NLevel0Cluster,
          index_t MLevel1Cluster,
          index_t NLevel1Cluster,
          index_t KPerThreadLoop>
Chao Liu's avatar
Chao Liu committed
228
229
230
231
struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
{
    struct MatrixIndex
    {
Chao Liu's avatar
Chao Liu committed
232
233
        index_t row;
        index_t col;
Chao Liu's avatar
Chao Liu committed
234
235
    };

Chao Liu's avatar
Chao Liu committed
236
237
    index_t mMyThreadOffsetA;
    index_t mMyThreadOffsetB;
Chao Liu's avatar
Chao Liu committed
238
239
240

    __device__ BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2()
    {
Chao Liu's avatar
Chao Liu committed
241
        constexpr index_t ThreadPerLevel1Cluster =
Chao Liu's avatar
Chao Liu committed
242
243
244
245
            MLevel0Cluster * NLevel0Cluster * MLevel1Cluster * NLevel1Cluster;

        static_assert(BlockSize == ThreadPerLevel1Cluster, "wrong! wrong blocksize\n");

Chao Liu's avatar
Chao Liu committed
246
247
248
        constexpr auto a_block_mtx  = BlockMatrixA{};
        constexpr auto b_block_mtx  = BlockMatrixB{};
        constexpr auto c_thread_mtx = ThreadMatrixC{};
Chao Liu's avatar
Chao Liu committed
249
250
251
252

        static_assert(a_block_mtx.NRow() == b_block_mtx.NRow(),
                      "wrong! K dimension not consistent\n");

Chao Liu's avatar
Chao Liu committed
253
254
255
        constexpr index_t M = a_block_mtx.NCol(); // A is transposed
        constexpr index_t N = b_block_mtx.NCol();
        constexpr index_t K = a_block_mtx.NRow();
Chao Liu's avatar
Chao Liu committed
256

Chao Liu's avatar
Chao Liu committed
257
258
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
259
260
261
262

        static_assert((MPerThread % MPerThreadSubC == 0) && (NPerThread % NPerThreadSubC == 0),
                      "wrong! Cannot evenly divide thread work among repeat \n");

Chao Liu's avatar
Chao Liu committed
263
264
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
265
266
267
268

        static_assert((M % MRepeat == 0) && (N % NRepeat == 0),
                      "wrong! Cannot evenly divide work among repeat\n");

Chao Liu's avatar
Chao Liu committed
269
270
        constexpr index_t MPerLevel1Cluster = M / MRepeat;
        constexpr index_t NPerLevel1Cluster = N / NRepeat;
Chao Liu's avatar
Chao Liu committed
271
272
273
274
275

        static_assert((MPerLevel1Cluster % MLevel1Cluster == 0) &&
                          (NPerLevel1Cluster % NLevel1Cluster == 0),
                      "wrong! Cannot evenly divide work among Level1Cluster\n");

Chao Liu's avatar
Chao Liu committed
276
277
        constexpr index_t MPerLevel0Cluster = MPerLevel1Cluster / MLevel1Cluster;
        constexpr index_t NPerLevel0Cluster = NPerLevel1Cluster / NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292

        static_assert((MPerLevel0Cluster % MLevel0Cluster == 0) &&
                          (NPerLevel0Cluster % NLevel0Cluster == 0),
                      "wrong! Cannot evenly divide work among Level0Cluster\n");

        static_assert((MPerThreadSubC == MPerLevel0Cluster / MLevel0Cluster) &&
                          (NPerThreadSubC == NPerLevel0Cluster / NLevel0Cluster),
                      "wrong! thread work size is wrong\n");

        auto c_thread_mtx_index = GetBeginOfThreadMatrixC(get_thread_local_1d_id());

        mMyThreadOffsetA = a_block_mtx.Get1dIndex(0, c_thread_mtx_index.row);
        mMyThreadOffsetB = b_block_mtx.Get1dIndex(0, c_thread_mtx_index.col);
    }

Chao Liu's avatar
Chao Liu committed
293
    __device__ static MatrixIndex GetBeginOfThreadMatrixC(index_t thread_id)
Chao Liu's avatar
Chao Liu committed
294
    {
Chao Liu's avatar
Chao Liu committed
295
        constexpr index_t ThreadPerLevel0Cluster = MLevel0Cluster * NLevel0Cluster;
Chao Liu's avatar
Chao Liu committed
296

Chao Liu's avatar
Chao Liu committed
297
298
299
        index_t level1_id   = thread_id / ThreadPerLevel0Cluster;
        index_t level1_m_id = level1_id / NLevel1Cluster;
        index_t level1_n_id = level1_id % NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
300

Chao Liu's avatar
Chao Liu committed
301
302
303
        index_t level0_id   = thread_id % ThreadPerLevel0Cluster;
        index_t level0_m_id = level0_id / NLevel0Cluster;
        index_t level0_n_id = level0_id % NLevel0Cluster;
Chao Liu's avatar
Chao Liu committed
304

Chao Liu's avatar
Chao Liu committed
305
306
        constexpr index_t MPerLevel0Cluster = MPerThreadSubC * MLevel0Cluster;
        constexpr index_t NPerLevel0Cluster = NPerThreadSubC * NLevel0Cluster;
Chao Liu's avatar
Chao Liu committed
307
308
309
310
311
312

        return MatrixIndex{level1_m_id * MPerLevel0Cluster + level0_m_id * MPerThreadSubC,
                           level1_n_id * NPerLevel0Cluster + level0_n_id * NPerThreadSubC};
    }

    // this should be optimized away if input is known
Chao Liu's avatar
Chao Liu committed
313
314
    __device__ static MatrixIndex GetDistanceFromBeginOfThreadMatrixC(index_t m_in_c,
                                                                      index_t n_in_c)
Chao Liu's avatar
Chao Liu committed
315
    {
Chao Liu's avatar
Chao Liu committed
316
        constexpr auto c_thread_mtx = ThreadMatrixC{};
Chao Liu's avatar
Chao Liu committed
317

Chao Liu's avatar
Chao Liu committed
318
319
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
320

Chao Liu's avatar
Chao Liu committed
321
322
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
323

Chao Liu's avatar
Chao Liu committed
324
325
        constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
        constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
326

Chao Liu's avatar
Chao Liu committed
327
328
        index_t m_repeat = m_in_c / MPerThreadSubC;
        index_t n_repeat = n_in_c / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
329

Chao Liu's avatar
Chao Liu committed
330
331
        index_t m_in_sub_c = m_in_c % MPerThreadSubC;
        index_t n_in_sub_c = n_in_c % NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
332
333
334
335
336
337

        return MatrixIndex{m_repeat * MPerLevel1Cluster + m_in_sub_c,
                           n_repeat * NPerLevel1Cluster + n_in_sub_c};
    }

    template <class FloatA, class FloatB, class FloatC, class Accumulator>
Chao Liu's avatar
Chao Liu committed
338
339
340
    __device__ void Run(const FloatA* __restrict__ p_a_block,
                        const FloatB* __restrict__ p_b_block,
                        FloatC* __restrict__ p_c_thread,
Chao Liu's avatar
Chao Liu committed
341
342
                        Accumulator f_accum) const
    {
Chao Liu's avatar
Chao Liu committed
343
344
        constexpr auto True  = integral_constant<bool, true>{};
        constexpr auto False = integral_constant<bool, false>{};
Chao Liu's avatar
Chao Liu committed
345

Chao Liu's avatar
Chao Liu committed
346
347
348
        constexpr auto a_block_mtx  = BlockMatrixA{};
        constexpr auto b_block_mtx  = BlockMatrixB{};
        constexpr auto c_thread_mtx = ThreadMatrixC{};
Chao Liu's avatar
Chao Liu committed
349

Chao Liu's avatar
Chao Liu committed
350
351
352
        constexpr index_t M = a_block_mtx.NCol();
        constexpr index_t N = b_block_mtx.NCol();
        constexpr index_t K = a_block_mtx.NRow();
Chao Liu's avatar
Chao Liu committed
353

Chao Liu's avatar
Chao Liu committed
354
355
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
356
357

        // thread A, B for GEMM
Chao Liu's avatar
Chao Liu committed
358
359
        constexpr auto a_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
Chao Liu's avatar
Chao Liu committed
360

Chao Liu's avatar
Chao Liu committed
361
362
        constexpr auto b_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
Chao Liu's avatar
Chao Liu committed
363
364

        // thread A-sub, B-sub for copy
Chao Liu's avatar
Chao Liu committed
365
366
        constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});
Chao Liu's avatar
Chao Liu committed
367

Chao Liu's avatar
Chao Liu committed
368
369
        constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
Chao Liu's avatar
Chao Liu committed
370

Jing Zhang's avatar
Jing Zhang committed
371
372
373
374
        float p_thread[a_thread_mtx.GetElementSpace() + b_thread_mtx.GetElementSpace()]; 

        FloatA *p_a_thread = p_thread;
        FloatB *p_b_thread = p_thread + a_thread_mtx.GetElementSpace();
Chao Liu's avatar
Chao Liu committed
375

Chao Liu's avatar
Chao Liu committed
376
377
        constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
        constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
378

Chao Liu's avatar
Chao Liu committed
379
380
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
381

382
#pragma unroll
Chao Liu's avatar
Chao Liu committed
383
        // loop over k
Chao Liu's avatar
Chao Liu committed
384
        for(index_t k_begin = 0; k_begin < K; k_begin += KPerThreadLoop)
Chao Liu's avatar
Chao Liu committed
385
        {
Jing Zhang's avatar
Jing Zhang committed
386
#if 0
Chao Liu's avatar
Chao Liu committed
387
            // copy A-sub to form A
Jing Zhang's avatar
Jing Zhang committed
388
389
390
#if 0
#pragma unroll
            // MRepeat = 2
Chao Liu's avatar
Chao Liu committed
391
            for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
Chao Liu's avatar
Chao Liu committed
392
            {
Chao Liu's avatar
Chao Liu committed
393
394
                threadwise_matrix_copy(
                    a_block_mtx,
Jing Zhang's avatar
inline  
Jing Zhang committed
395
                    //MPerLevel1Cluster = 4
Chao Liu's avatar
Chao Liu committed
396
397
398
                    p_a_block + a_block_mtx.Get1dIndex(k_begin, m_repeat * MPerLevel1Cluster) +
                        mMyThreadOffsetA,
                    a_thread_mtx,
Jing Zhang's avatar
inline  
Jing Zhang committed
399
                    //MPerThreadSubC = 4
Chao Liu's avatar
Chao Liu committed
400
401
                    p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
                    a_thread_sub_mtx.GetLengths());
Chao Liu's avatar
Chao Liu committed
402
            }
Jing Zhang's avatar
Jing Zhang committed
403
404
405
406
#else
            {
                auto src_index = a_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetA;
                auto dst_index = a_thread_sub_mtx.Get1dIndex(0, 0);
Chao Liu's avatar
Chao Liu committed
407

Jing Zhang's avatar
Jing Zhang committed
408
                const float4* loc = (const float4 *)(p_a_block + src_index);
Jing Zhang's avatar
Jing Zhang committed
409
410
                float4* reg = (float4 *)(p_a_thread + dst_index);

Jing Zhang's avatar
Jing Zhang committed
411
                reg[0] = loc[0];
Jing Zhang's avatar
Jing Zhang committed
412
413
                reg[1] = loc[16];
                //reg[MPerThreadSubC/4] = loc[MPerLevel1Cluster/4];
Jing Zhang's avatar
Jing Zhang committed
414
415
                //asm volatile("\n \
                        //ds_read2_b64 %0, %2 offset1:1 \n \
Jing Zhang's avatar
Jing Zhang committed
416
                        //ds_read2_b64 %1, %2 offset0:32 offset1:33 \n \
Jing Zhang's avatar
Jing Zhang committed
417
                        //s_waitcnt lgkmcnt(0)" 
Jing Zhang's avatar
Jing Zhang committed
418
419
                        //: "=v"(reg[0]), "=v"(reg[1]) 
                        //: "v"(__to_local((void *)(loc)))
Jing Zhang's avatar
Jing Zhang committed
420
                        //);
Jing Zhang's avatar
Jing Zhang committed
421
422
423
424
            }
#endif

#if 0
Chao Liu's avatar
Chao Liu committed
425
            // copy B-sub to form B
Jing Zhang's avatar
Jing Zhang committed
426
#pragma unroll
Chao Liu's avatar
Chao Liu committed
427
            for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
Chao Liu's avatar
Chao Liu committed
428
            {
Chao Liu's avatar
Chao Liu committed
429
430
431
432
433
434
435
                threadwise_matrix_copy(
                    b_block_mtx,
                    p_b_block + b_block_mtx.Get1dIndex(k_begin, n_repeat * NPerLevel1Cluster) +
                        mMyThreadOffsetB,
                    b_thread_mtx,
                    p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                    b_thread_sub_mtx.GetLengths());
Chao Liu's avatar
Chao Liu committed
436
            }
Jing Zhang's avatar
Jing Zhang committed
437
438
439
440
441
442
443
444
445
#else
            {
                auto src_index = b_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetB;
                auto dst_index = b_thread_sub_mtx.Get1dIndex(0, 0);

                const float4* loc = (const float4 *)(p_b_block + src_index);
                float4* reg = (float4 *)(p_b_thread + dst_index);

                reg[0] = loc[0];
Jing Zhang's avatar
Jing Zhang committed
446
447
448
449
450
451
452
453
454
                reg[1] = loc[8];
                //reg[NPerThreadSubC/4] = loc[NPerLevel1Cluster/4];
                //asm volatile("\n \
                        //ds_read2_b64 %0, %2 offset1:1 \n \
                        //ds_read2_b64 %1, %2 offset0:16 offset1:17 \n \
                        //s_waitcnt lgkmcnt(0)" 
                        //: "=v"(reg[0]), "=v"(reg[1]) 
                        //: "v"(__to_local((void *)(loc)))
                        //);
Jing Zhang's avatar
Jing Zhang committed
455
            }
Jing Zhang's avatar
Jing Zhang committed
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
#endif

#else
            auto a_src_index = a_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetA;
            auto b_src_index = b_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetB;
            auto dst_index = a_thread_sub_mtx.Get1dIndex(0, 0);

            const float4* a_loc = (const float4 *)(p_a_block + a_src_index);
            const float4* b_loc = (const float4 *)(p_b_block + b_src_index);
            float4* reg = (float4 *)(p_a_thread + dst_index);

            //reg[0] = a_loc[0];
            //reg[1] = a_loc[16];
            //reg[2] = b_loc[0];
            //reg[3] = b_loc[8];
            //s_waitcnt lgkmcnt(0)                                       // 000000001398: BF8CC07F 
            asm volatile("\n \
                    ds_read2_b64 %0, %4 offset1:1 \n \
                    ds_read2_b64 %1, %4 offset0:32 offset1:33 \n \
                    ds_read2_b64 %2, %5 offset1:1 \n \
                    ds_read2_b64 %3, %5 offset0:16 offset1:17 \n \
                    s_waitcnt lgkmcnt(0)" 
                    : "=v"(reg[0]), "=v"(reg[1]), "=v"(reg[2]), "=v"(reg[3]) 
                    : "v"(__to_local((void *)(a_loc))), "v"(__to_local((void *)(b_loc)))
                    );


Jing Zhang's avatar
Jing Zhang committed
483
#endif
Chao Liu's avatar
Chao Liu committed
484
485

            // C = A * B
Jing Zhang's avatar
Jing Zhang committed
486
#if 0
Chao Liu's avatar
Chao Liu committed
487
488
489
490
491
492
493
494
495
496
            threadwise_gemm(a_thread_mtx,
                            True,
                            p_a_thread,
                            b_thread_mtx,
                            False,
                            p_b_thread,
                            c_thread_mtx,
                            False,
                            p_c_thread,
                            f_accum);
Jing Zhang's avatar
Jing Zhang committed
497
498
499
500
#else
            for(index_t k = 0; k < 1; ++k)
            {
                // M = 8
Jing Zhang's avatar
Jing Zhang committed
501
                const index_t bindex = b_thread_sub_mtx.Get1dIndex(k, 0);
Jing Zhang's avatar
Jing Zhang committed
502
503
504
                for(index_t i = 0; i < 8; ++i)
                {
                    // N = 8
Jing Zhang's avatar
Jing Zhang committed
505
506
507
                    const index_t aindex = a_thread_sub_mtx.Get1dIndex(k, i); // A is transposed
                    const index_t cindex = c_thread_mtx.Get1dIndex(i, 0);
                    //for(index_t j = 0; j < 8; ++j)
Jing Zhang's avatar
Jing Zhang committed
508
509
                    {

Jing Zhang's avatar
Jing Zhang committed
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
                        //p_c_thread[cindex] += p_a_thread[aindex] * p_b_thread[bindex];
                        asm volatile("\n \
                                v_mac_f32 %0, %8, %9 \n \
                                v_mac_f32 %1, %8, %10 \n \
                                v_mac_f32 %2, %8, %11 \n \
                                v_mac_f32 %3, %8, %12 \n \
                                v_mac_f32 %4, %8, %13 \n \
                                v_mac_f32 %5, %8, %14 \n \
                                v_mac_f32 %6, %8, %15 \n \
                                v_mac_f32 %7, %8, %16 \n \
                                "
                                : "=v"(p_c_thread[cindex + 0]),
                                  "=v"(p_c_thread[cindex + 1]),
                                  "=v"(p_c_thread[cindex + 2]),
                                  "=v"(p_c_thread[cindex + 3]),
                                  "=v"(p_c_thread[cindex + 4]),
                                  "=v"(p_c_thread[cindex + 5]),
                                  "=v"(p_c_thread[cindex + 6]),
                                  "=v"(p_c_thread[cindex + 7])
                                : "v"(p_a_thread[aindex]), 
                                  "v"(p_b_thread[bindex + 0]), 
                                  "v"(p_b_thread[bindex + 1]),
                                  "v"(p_b_thread[bindex + 2]),
                                  "v"(p_b_thread[bindex + 3]),
                                  "v"(p_b_thread[bindex + 4]),
                                  "v"(p_b_thread[bindex + 5]),
                                  "v"(p_b_thread[bindex + 6]),
Jing Zhang's avatar
Jing Zhang committed
537
                                  "v"(p_b_thread[bindex + 7]),
Jing Zhang's avatar
Jing Zhang committed
538
539
540
541
542
543
544
545
546
                                  "0"(p_c_thread[cindex + 0]),
                                  "1"(p_c_thread[cindex + 1]),
                                  "2"(p_c_thread[cindex + 2]),
                                  "3"(p_c_thread[cindex + 3]),
                                  "4"(p_c_thread[cindex + 4]),
                                  "5"(p_c_thread[cindex + 5]),
                                  "6"(p_c_thread[cindex + 6]),
                                  "7"(p_c_thread[cindex + 7])
                                );
Jing Zhang's avatar
Jing Zhang committed
547
548
549
550
                    }
                }
            }
#endif
Chao Liu's avatar
Chao Liu committed
551
552
        }
    }
553

Chao Liu's avatar
Chao Liu committed
554
    template <class FloatA, class FloatB, class FloatC, class Accumulator>
Chao Liu's avatar
Chao Liu committed
555
556
557
    __device__ void Run_asm(const FloatA* const __restrict__ p_a_block,
                            const FloatB* const __restrict__ p_b_block,
                            FloatC* const __restrict__ p_c_thread,
Chao Liu's avatar
Chao Liu committed
558
559
560
561
562
563
564
565
566
                            Accumulator f_accum) const
    {
        constexpr auto True  = integral_constant<bool, true>{};
        constexpr auto False = integral_constant<bool, false>{};

        constexpr auto a_block_mtx  = BlockMatrixA{};
        constexpr auto b_block_mtx  = BlockMatrixB{};
        constexpr auto c_thread_mtx = ThreadMatrixC{};

Chao Liu's avatar
Chao Liu committed
567
568
569
        constexpr index_t M = a_block_mtx.NCol();
        constexpr index_t N = b_block_mtx.NCol();
        constexpr index_t K = a_block_mtx.NRow();
Chao Liu's avatar
Chao Liu committed
570

Chao Liu's avatar
Chao Liu committed
571
572
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590

        // thread A, B for GEMM
        constexpr auto a_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});

        constexpr auto b_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});

        // thread A-sub, B-sub for copy
        constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});

        constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});

        FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
        FloatB p_b_thread[b_thread_mtx.GetElementSpace()];

Chao Liu's avatar
Chao Liu committed
591
592
        constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
        constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
593

Chao Liu's avatar
Chao Liu committed
594
595
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
596

Chao Liu's avatar
Chao Liu committed
597
598
599
600
601
602
        static_assert(MPerThreadSubC == 4 && NPerThreadSubC == 4 && MRepeat == 2 && NRepeat == 2 &&
                          KPerThreadLoop == 1 && K == 1,
                      "asm is not for this mtx shape");

        const FloatA* const p_a_block_thread_offset = p_a_block + mMyThreadOffsetA;

Chao Liu's avatar
Chao Liu committed
603
604
#pragma unroll
        // loop over k
Chao Liu's avatar
Chao Liu committed
605
        for(index_t k_begin = 0; k_begin < K; k_begin += KPerThreadLoop)
Chao Liu's avatar
Chao Liu committed
606
        {
Chao Liu's avatar
Chao Liu committed
607
608
#if 0
#pragma unroll
Chao Liu's avatar
Chao Liu committed
609
            // copy A-sub to form A
Chao Liu's avatar
Chao Liu committed
610
            for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
Chao Liu's avatar
Chao Liu committed
611
612
613
614
615
616
            {
                threadwise_matrix_copy(
                    a_block_mtx,
                    p_a_block + a_block_mtx.Get1dIndex(k_begin, m_repeat * MPerLevel1Cluster) +
                        mMyThreadOffsetA,
                    a_thread_mtx,
Chao Liu's avatar
Chao Liu committed
617
                    a_thread_sub_mtx.NCol(p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
Chao Liu's avatar
Chao Liu committed
618
619
                    a_thread_sub_mtx.GetLengths());
            }
Chao Liu's avatar
Chao Liu committed
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
#elif 1
            // this produce right result
            using vectorA_t = typename vector_type<FloatA, 4>::MemoryType; // this is float4*

            asm volatile(
                "\n \
                    ds_read_b128 %0, %1 \n \
                    s_waitcnt lgkmcnt(0)"
                : "=v"(*(reinterpret_cast<vectorA_t*>(p_a_thread + a_thread_mtx.Get1dIndex(0, 0))))
                : "v"(__to_local(
                    (void*)(p_a_block + a_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetA))));

            asm volatile("\n \
                    ds_read_b128 %0, %1 \n \
                    s_waitcnt lgkmcnt(0)"
                         : "=v"(*(reinterpret_cast<vectorA_t*>(
                             p_a_thread + a_thread_mtx.Get1dIndex(0, MPerThreadSubC))))
                         : "v"(__to_local((
                             void*)(p_a_block + a_block_mtx.Get1dIndex(k_begin, MPerLevel1Cluster) +
                                    mMyThreadOffsetA))));
#elif 0
            // this produce wrong result
            using vectorA_t = typename vector_type<FloatA, 4>::MemoryType; // this is float4*

            asm volatile(
                "\n \
                    ds_read_b128 %0, %2  \n \
                    ds_read_b128 %1, %3  \n \
                    s_waitcnt lgkmcnt(0)"
                : "=v"(*(reinterpret_cast<vectorA_t*>(p_a_thread + a_thread_mtx.Get1dIndex(0, 0)))),
                  "=v"(*(reinterpret_cast<vectorA_t*>(p_a_thread +
                                                      a_thread_mtx.Get1dIndex(0, MPerThreadSubC))))
                : "v"(__to_local(
                      (void*)(p_a_block + a_block_mtx.Get1dIndex(k_begin, 0) + mMyThreadOffsetA))),
                  "v"(__to_local((void*)(p_a_block +
                                         a_block_mtx.Get1dIndex(k_begin, MPerLevel1Cluster) +
                                         mMyThreadOffsetA))));
#elif 1
            // this produce wrong result
            using vectorA_t = typename vector_type<FloatA, 4>::MemoryType; // this is float4*

            asm volatile(
                "\n \
                    ds_read_b128 %0, %1 \n \
                    s_waitcnt lgkmcnt(0)"
                : "=v"(*(reinterpret_cast<vectorA_t*>(p_a_thread + a_thread_mtx.Get1dIndex(0, 0))))
                : "v"(__to_local((void*)(p_a_block_thread_offset))));

            asm volatile("\n \
                    ds_read_b128 %0, %1 offset:16 \n \
                    s_waitcnt lgkmcnt(0)"
                         : "=v"(*(reinterpret_cast<vectorA_t*>(
                             p_a_thread + a_thread_mtx.Get1dIndex(0, MPerThreadSubC))))
                         : "v"(__to_local((void*)(p_a_block_thread_offset))));

#endif
Chao Liu's avatar
Chao Liu committed
676

Chao Liu's avatar
Chao Liu committed
677
            //#pragma unroll
Chao Liu's avatar
Chao Liu committed
678
            // copy B-sub to form B
Chao Liu's avatar
Chao Liu committed
679
            for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
Chao Liu's avatar
Chao Liu committed
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
            {
                threadwise_matrix_copy(
                    b_block_mtx,
                    p_b_block + b_block_mtx.Get1dIndex(k_begin, n_repeat * NPerLevel1Cluster) +
                        mMyThreadOffsetB,
                    b_thread_mtx,
                    p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                    b_thread_sub_mtx.GetLengths());
            }

// C = A * B
#if 1
            threadwise_gemm(a_thread_mtx,
                            True,
                            p_a_thread,
                            b_thread_mtx,
                            False,
                            p_b_thread,
                            c_thread_mtx,
                            False,
                            p_c_thread,
                            f_accum);
Chao Liu's avatar
Chao Liu committed
702
#elif 0
Chao Liu's avatar
Chao Liu committed
703
704
705
706
            // inline asm
            static_assert(c_thread_mtx.NRow() == 8 && c_thread_mtx.NCol() == 8,
                          "asm is only for 8x8");

Chao Liu's avatar
Chao Liu committed
707
            for(index_t k = 0; k < a_thread_mtx.NRow(); ++k) // A is transposed
Chao Liu's avatar
Chao Liu committed
708
            {
Chao Liu's avatar
Chao Liu committed
709
                const index_t bindex = b_thread_mtx.Get1dIndex(k, 0);
Chao Liu's avatar
Chao Liu committed
710

Chao Liu's avatar
Chao Liu committed
711
                for(index_t i = 0; i < c_thread_mtx.NRow(); ++i)
Chao Liu's avatar
Chao Liu committed
712
                {
Chao Liu's avatar
Chao Liu committed
713
714
                    const index_t aindex = a_thread_mtx.Get1dIndex(k, i); // A is transposed
                    const index_t cindex = c_thread_mtx.Get1dIndex(i, 0);
Chao Liu's avatar
Chao Liu committed
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756

                    asm volatile("\n \
                                v_mac_f32 %0, %8, %9 \n \
                                v_mac_f32 %1, %8, %10 \n \
                                v_mac_f32 %2, %8, %11 \n \
                                v_mac_f32 %3, %8, %12 \n \
                                v_mac_f32 %4, %8, %13 \n \
                                v_mac_f32 %5, %8, %14 \n \
                                v_mac_f32 %6, %8, %15 \n \
                                v_mac_f32 %7, %8, %16 \n \
                                "
                                 : "=v"(p_c_thread[cindex + 0]),
                                   "=v"(p_c_thread[cindex + 1]),
                                   "=v"(p_c_thread[cindex + 2]),
                                   "=v"(p_c_thread[cindex + 3]),
                                   "=v"(p_c_thread[cindex + 4]),
                                   "=v"(p_c_thread[cindex + 5]),
                                   "=v"(p_c_thread[cindex + 6]),
                                   "=v"(p_c_thread[cindex + 7])
                                 : "v"(p_a_thread[aindex]),
                                   "v"(p_b_thread[bindex + 0]),
                                   "v"(p_b_thread[bindex + 1]),
                                   "v"(p_b_thread[bindex + 2]),
                                   "v"(p_b_thread[bindex + 3]),
                                   "v"(p_b_thread[bindex + 4]),
                                   "v"(p_b_thread[bindex + 5]),
                                   "v"(p_b_thread[bindex + 6]),
                                   "v"(p_b_thread[bindex + 7]),
                                   "0"(p_c_thread[cindex + 0]),
                                   "1"(p_c_thread[cindex + 1]),
                                   "2"(p_c_thread[cindex + 2]),
                                   "3"(p_c_thread[cindex + 3]),
                                   "4"(p_c_thread[cindex + 4]),
                                   "5"(p_c_thread[cindex + 5]),
                                   "6"(p_c_thread[cindex + 6]),
                                   "7"(p_c_thread[cindex + 7]));
                }
            }
#endif
        }
    }

757
758
759
760
761
762
    template <class FloatA, class FloatB, class FloatC, class Accumulator>
    __device__ void Run_RegisterDoubleBuffer(FloatA* const p_a_block,
                                             FloatB* const p_b_block,
                                             FloatC* p_c_thread,
                                             Accumulator f_accum) const
    {
Chao Liu's avatar
Chao Liu committed
763
764
        constexpr auto True  = integral_constant<bool, true>{};
        constexpr auto False = integral_constant<bool, false>{};
765

Chao Liu's avatar
Chao Liu committed
766
767
768
        constexpr auto a_block_mtx  = BlockMatrixA{};
        constexpr auto b_block_mtx  = BlockMatrixB{};
        constexpr auto c_thread_mtx = ThreadMatrixC{};
769

Chao Liu's avatar
Chao Liu committed
770
771
772
        constexpr index_t M = a_block_mtx.NCol();
        constexpr index_t N = b_block_mtx.NCol();
        constexpr index_t K = a_block_mtx.NRow();
773

Chao Liu's avatar
Chao Liu committed
774
775
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
776
777

        // thread A, B for GEMM
Chao Liu's avatar
Chao Liu committed
778
779
        constexpr auto a_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
780

Chao Liu's avatar
Chao Liu committed
781
782
        constexpr auto b_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
783
784

        // thread A-sub, B-sub for copy
Chao Liu's avatar
Chao Liu committed
785
786
        constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});
787

Chao Liu's avatar
Chao Liu committed
788
789
        constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
790

791
        // register
792
793
794
795
796
797
        FloatA p_a_thread_0[a_thread_mtx.GetElementSpace()];
        FloatB p_b_thread_0[b_thread_mtx.GetElementSpace()];

        FloatA p_a_thread_1[a_thread_mtx.GetElementSpace()];
        FloatB p_b_thread_1[b_thread_mtx.GetElementSpace()];

Chao Liu's avatar
Chao Liu committed
798
799
        constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
        constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
800

Chao Liu's avatar
Chao Liu committed
801
802
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
803

Chao Liu's avatar
Chao Liu committed
804
// preload A, B
805
#pragma unroll
Chao Liu's avatar
Chao Liu committed
806
        for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
807
808
809
810
811
812
813
814
815
        { // copy A-sub to form A
            threadwise_matrix_copy(a_block_mtx,
                                   p_a_block + mMyThreadOffsetA + m_repeat * MPerLevel1Cluster,
                                   a_thread_sub_mtx,
                                   p_a_thread_0 + m_repeat * MPerThreadSubC,
                                   a_thread_sub_mtx.GetLengths());
        }

#pragma unroll
Chao Liu's avatar
Chao Liu committed
816
        for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
817
818
819
820
821
822
823
824
825
826
827
        { // copy B-sub to form B
            threadwise_matrix_copy(b_block_mtx,
                                   p_b_block + mMyThreadOffsetB + n_repeat * NPerLevel1Cluster,
                                   b_thread_sub_mtx,
                                   p_b_thread_0 + n_repeat * NPerThreadSubC,
                                   b_thread_sub_mtx.GetLengths());
        }

        bool even_loop = true;

#pragma unroll
Chao Liu's avatar
Chao Liu committed
828
        for(index_t k_begin = 0; k_begin + KPerThreadLoop < K;
829
830
831
832
833
834
835
836
            k_begin += KPerThreadLoop, even_loop = !even_loop)
        { // loop over k
            FloatA* p_a_thread_now = even_loop ? p_a_thread_0 : p_a_thread_1;
            FloatB* p_b_thread_now = even_loop ? p_b_thread_0 : p_b_thread_1;

            FloatA* p_a_thread_next = even_loop ? p_a_thread_1 : p_a_thread_0;
            FloatB* p_b_thread_next = even_loop ? p_b_thread_1 : p_b_thread_0;

Chao Liu's avatar
Chao Liu committed
837
// preload next A, B
838
#pragma unroll
Chao Liu's avatar
Chao Liu committed
839
            for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
840
841
842
843
844
845
846
847
848
849
850
            { // copy A-sub to form A
                threadwise_matrix_copy(a_block_mtx,
                                       p_a_block + mMyThreadOffsetA +
                                           (k_begin + 1) * a_block_mtx.RowStride() +
                                           m_repeat * MPerLevel1Cluster,
                                       a_thread_sub_mtx,
                                       p_a_thread_next + m_repeat * MPerThreadSubC,
                                       a_thread_sub_mtx.GetLengths());
            }

#pragma unroll
Chao Liu's avatar
Chao Liu committed
851
            for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
            { // copy B-sub to form B
                threadwise_matrix_copy(b_block_mtx,
                                       p_b_block + mMyThreadOffsetB +
                                           (k_begin + 1) * b_block_mtx.RowStride() +
                                           n_repeat * NPerLevel1Cluster,
                                       b_thread_sub_mtx,
                                       p_b_thread_next + n_repeat * NPerThreadSubC,
                                       b_thread_sub_mtx.GetLengths());
            }

            // C = A * B
            threadwise_gemm(a_thread_mtx,
                            True,
                            p_a_thread_now,
                            b_thread_mtx,
                            False,
                            p_b_thread_now,
                            c_thread_mtx,
                            False,
                            p_c_thread,
                            f_accum);
        }

        // last loop
        {
            FloatA* p_a_thread_now = even_loop ? p_a_thread_0 : p_a_thread_1;
            FloatB* p_b_thread_now = even_loop ? p_b_thread_0 : p_b_thread_1;

            // C = A * B
            threadwise_gemm(a_thread_mtx,
                            True,
                            p_a_thread_now,
                            b_thread_mtx,
                            False,
                            p_b_thread_now,
                            c_thread_mtx,
                            False,
                            p_c_thread,
                            f_accum);
        }
    }
Chao Liu's avatar
Chao Liu committed
893
894
895
896
897
898
899
900
901
902
903
904
905
906

    template <class FloatA, class FloatB, class FloatC, class Accumulator>
    __device__ void Run_v2(const FloatA* __restrict__ p_a_block,
                           const FloatB* __restrict__ p_b_block,
                           FloatC* __restrict__ p_c_thread,
                           Accumulator f_accum) const
    {
        constexpr auto True  = integral_constant<bool, true>{};
        constexpr auto False = integral_constant<bool, false>{};

        constexpr auto a_block_mtx  = BlockMatrixA{};
        constexpr auto b_block_mtx  = BlockMatrixB{};
        constexpr auto c_thread_mtx = ThreadMatrixC{};

Chao Liu's avatar
Chao Liu committed
907
908
909
        constexpr index_t M = a_block_mtx.NCol();
        constexpr index_t N = b_block_mtx.NCol();
        constexpr index_t K = a_block_mtx.NRow();
Chao Liu's avatar
Chao Liu committed
910

Chao Liu's avatar
Chao Liu committed
911
912
        constexpr index_t MPerThread = c_thread_mtx.NRow();
        constexpr index_t NPerThread = c_thread_mtx.NCol();
Chao Liu's avatar
Chao Liu committed
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933

        // thread A-sub, B-sub, C-sub
        constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});

        constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});

        constexpr auto c_thread_sub_mtx = make_ConstantMatrixDescriptor(
            Number<MPerThreadSubC>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});

        // thread A, B
        constexpr auto a_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});

        constexpr auto b_thread_mtx =
            make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});

        FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
        FloatB p_b_thread[b_thread_mtx.GetElementSpace()];

Chao Liu's avatar
Chao Liu committed
934
935
        constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
        constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
Chao Liu's avatar
Chao Liu committed
936

Chao Liu's avatar
Chao Liu committed
937
938
        constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
        constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
Chao Liu's avatar
Chao Liu committed
939
940
941

#pragma unroll
        // loop over k
Chao Liu's avatar
Chao Liu committed
942
        for(index_t k_begin = 0; k_begin < K; k_begin += KPerThreadLoop)
Chao Liu's avatar
Chao Liu committed
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
        {
            // C-sub(s) in first row-wise subblock of C
            {
                //   copy first A-sub
                threadwise_matrix_copy(a_block_mtx,
                                       p_a_block + a_block_mtx.Get1dIndex(k_begin, 0) +
                                           mMyThreadOffsetA,
                                       a_thread_mtx,
                                       p_a_thread,
                                       a_thread_sub_mtx.GetLengths());

                //   copy first B-sub
                threadwise_matrix_copy(b_block_mtx,
                                       p_b_block + b_block_mtx.Get1dIndex(k_begin, 0) +
                                           mMyThreadOffsetB,
                                       b_thread_mtx,
                                       p_b_thread,
                                       b_thread_sub_mtx.GetLengths());

                //   do first sub GEMM
                threadwise_gemm(a_thread_sub_mtx,
                                True,
                                p_a_thread,
                                b_thread_sub_mtx,
                                False,
                                p_b_thread,
                                c_thread_sub_mtx,
                                False,
                                p_c_thread,
                                f_accum);

#pragma unroll
                //   copy next B-sub, and do GEMM
Chao Liu's avatar
Chao Liu committed
976
                for(index_t n_repeat = 1; n_repeat < NRepeat; ++n_repeat)
Chao Liu's avatar
Chao Liu committed
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
                {
                    threadwise_matrix_copy(
                        b_block_mtx,
                        p_b_block + b_block_mtx.Get1dIndex(k_begin, n_repeat * NPerLevel1Cluster) +
                            mMyThreadOffsetB,
                        b_thread_mtx,
                        p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                        b_thread_sub_mtx.GetLengths());

                    threadwise_gemm(
                        a_thread_sub_mtx,
                        True,
                        p_a_thread,
                        b_thread_sub_mtx,
                        False,
                        p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                        c_thread_sub_mtx,
                        False,
                        p_c_thread + c_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                        f_accum);
                }

#pragma unroll
                // loop over rest of row-wise subblock
                //   all B-sub(s) has been copied, so only A-sub(s) need to be copied
Chao Liu's avatar
Chao Liu committed
1002
                for(index_t m_repeat = 1; m_repeat < MRepeat; ++m_repeat)
Chao Liu's avatar
Chao Liu committed
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
                {
                    // copy a A-sub
                    threadwise_matrix_copy(
                        a_block_mtx,
                        p_a_block + a_block_mtx.Get1dIndex(k_begin, m_repeat * MPerLevel1Cluster) +
                            mMyThreadOffsetA,
                        a_thread_mtx,
                        p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
                        a_thread_sub_mtx.GetLengths());

                    // do some GEMMs
Chao Liu's avatar
Chao Liu committed
1014
                    for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
Chao Liu's avatar
Chao Liu committed
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
                    {
                        threadwise_gemm(
                            a_thread_sub_mtx,
                            True,
                            p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
                            b_thread_sub_mtx,
                            False,
                            p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                            c_thread_sub_mtx,
                            False,
Chao Liu's avatar
Chao Liu committed
1025
1026
1027
                            p_c_thread +
                                c_thread_mtx.Get1dIndex(m_repeat * MPerThreadSubC,
                                                        n_repeat * NPerThreadSubC),
Chao Liu's avatar
Chao Liu committed
1028
1029
1030
1031
1032
1033
                            f_accum);
                    }
                }
            }
        }
    }
Chao Liu's avatar
Chao Liu committed
1034
};