Blame - src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl - ml/ComputeLibrary

* @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3

116

* @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform

117

* @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2

118

* @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2

119

* @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3

120

*

121

* All parameters are similarly defined in kernel gemm_mm_native, with these additions:

122

*

123

* @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32

124

* @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes)

125

* @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes)

126

* @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes)

127

* @param[in] eltwise_operand_step_y eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes)

128

* @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes)

129

*/

130

__kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),

131

IMAGE_DECLARATION(rhs),

132

#if defined(BETA)

133

IMAGE_DECLARATION(bias),

134

#endif // defined(BETA)

135

IMAGE_DECLARATION(dst),

SiCongLi

d928735

2021-11-03 19:01:22 +0000

[diff] [blame^]

136

// Post Op arguments

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

137

IMAGE_DECLARATION(eltwise_operand),

uint lhs_stride_z,

uint rhs_stride_z,

#if defined(BETA)

uint bias_stride_z,

#endif //defined(BETA)

143

uint dst_stride_z,

144

uint eltwise_operand_stride_z

145

#if defined(REINTERPRET_INPUT_AS_3D)

146

,

147

uint lhs_cross_plane_pad

148

#endif // REINTERPRET_INPUT_AS_3D

149

#if defined(REINTERPRET_OUTPUT_AS_3D)

150

,

151

uint dst_cross_plane_pad

152

#endif // REINTERPRET_OUTPUT_AS_3D

)

{

// Block size

#define RHS_BLOCK_SIZE ((K0) * (N0))

157

158

// RHS offset and step X

159

#define RHS_OFFSET_X (RHS_BLOCK_SIZE)

160

161

uint x = get_global_id(0);

162

uint y = get_global_id(1);

163

uint z = get_global_id(2);

164

165

#if defined(DUMMY_WORK_ITEMS)

166

if((x * N0 >= N) || (y * M0 >= M))

{

return;

}

#endif // defined(DUMMY_WORK_ITEMS)

171

172

// Compute LHS matrix address

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

173

uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

174

175

// Compute RHS matrix address

176

uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);

177

178

#if defined(MATRIX_B_DEPTH)

179

// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3

180

rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;

181

#else // defined(MATRIX_B_DEPTH)

182

rhs_offset += z * rhs_stride_z;

183

#endif // defined(MATRIX_B_DEPTH)

184

185

REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);

186

REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

187

188

#if defined(REINTERPRET_INPUT_AS_3D)

189

// The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

190

CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

191

192

// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we

193

// multiply lhs_stride_z by DEPTH_GEMM3D

194

lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

195

196

#else // defined(REINTERPRET_INPUT_AS_3D)

197

198

// Add offset for batched GEMM

199

lhs_offset += z * lhs_stride_z;

200

201

#endif // defined(REINTERPRET_INPUT_AS_3D)

202

203

// Initialize the accumulators

204

REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0;

int i = 0;

#if K0 > 1

for(; i <= (K - K0); i += K0)

209

{

210

// Supported cases (M0, K0):

211

// 1,2 - 1,3 - 1,4 - 1,8 - 1,16

212

// 2,2 - 2,3 - 2,4 - 2,8 - 2,16

213

// 3,2 - 3,3 - 3,4 - 3,8 - 3,16

214

// 4,2 - 4,3 - 4,4 - 4,8 - 4,16

215

// 5,2 - 5,3 - 5,4 - 5,8 - 5,16

216

// 6,2 - 6,3 - 6,4 - 6,8 - 6,16

217

// 7,2 - 7,3 - 7,4 - 7,8 - 7,16

218

// 8,2 - 8,3 - 8,4 - 8,8 - 8,16

219

// Load values from LHS matrix

220

LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);

221

222

// Load values from RHS matrix

223

LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero);

224

225

RHS_VFMA_M0xN0(0, a, b0, c);

226

RHS_VFMA_M0xN0(1, a, b1, c);

227

#if K0 > 2

228

RHS_VFMA_M0xN0(2, a, b2, c);

229

#endif // K0 > 2

230

#if K0 > 3

231

RHS_VFMA_M0xN0(3, a, b3, c);

232

#endif // K0 > 3

233

#if K0 > 4

234

RHS_VFMA_M0xN0(4, a, b4, c);

235

RHS_VFMA_M0xN0(5, a, b5, c);

236

RHS_VFMA_M0xN0(6, a, b6, c);

237

RHS_VFMA_M0xN0(7, a, b7, c);

238

#endif // K0 > 4

239

#if K0 > 8

240

RHS_VFMA_M0xN0(8, a, b8, c);

241

RHS_VFMA_M0xN0(9, a, b9, c);

242

RHS_VFMA_M0xN0(A, a, bA, c);

243

RHS_VFMA_M0xN0(B, a, bB, c);

244

RHS_VFMA_M0xN0(C, a, bC, c);

245

RHS_VFMA_M0xN0(D, a, bD, c);

246

RHS_VFMA_M0xN0(E, a, bE, c);

247

RHS_VFMA_M0xN0(F, a, bF, c);

248

#endif // K0 > 8

249

250

lhs_offset += K0 * sizeof(DATA_TYPE);

251

rhs_offset += K0 * rhs_stride_y;

252

}

253

#endif // K0 > 1

254

// Left-over accumulations

255

for(; i < K; ++i)

256

{

257

// Load values from LHS matrix

258

VEC_DATA_TYPE(DATA_TYPE, 2)

259

a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0));

260

#if M0 > 1

261

VEC_DATA_TYPE(DATA_TYPE, 2)

262

a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1));

263

#endif // M0 > 1

264

#if M0 > 2

265

VEC_DATA_TYPE(DATA_TYPE, 2)

266

a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2));

267

#endif // M0 > 2

268

#if M0 > 3

269

VEC_DATA_TYPE(DATA_TYPE, 2)

270

a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3));

271

#endif // M0 > 3

272

#if M0 > 4

273

VEC_DATA_TYPE(DATA_TYPE, 2)

274

a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4));

275

#endif // M0 > 4

276

#if M0 > 5

277

VEC_DATA_TYPE(DATA_TYPE, 2)

278

a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5));

279

#endif // M0 > 5

280

#if M0 > 6

281

VEC_DATA_TYPE(DATA_TYPE, 2)

282

a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6));

283

#endif // M0 > 6

284

#if M0 > 7

285

VEC_DATA_TYPE(DATA_TYPE, 2)

286

a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7));

287

#endif // M0 > 7

288

289

VEC_DATA_TYPE(DATA_TYPE, N0)

290

b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y));

291

RHS_VFMA_M0xN0(0, a, b, c);

292

293

lhs_offset += sizeof(DATA_TYPE);

294

rhs_offset += rhs_stride_y;

295

}

296

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

297

__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

298

299

REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

300

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

301

#if defined(REINTERPRET_OUTPUT_AS_3D)

302

// The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

303

CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

304

305

// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we

306

// multiply dst_stride_z by DEPTH_GEMM3D

307

dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

308

309

#else // defined(REINTERPRET_OUTPUT_AS_3D)

310

311

// Add offset for batched GEMM

312

dst_addr += z * dst_stride_z;

313

314

#endif // defined(REINTERPRET_OUTPUT_AS_3D)

315

316

// Multiply by the weight of matrix-matrix product and store the result

317

#if defined(ALPHA)

318

SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);

319

#endif // defined(ALPHA)

// Add beta*bias

#if defined(BETA)

#if defined(BROADCAST_BIAS)

324

__global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

325

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

326

LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

327

328

#ifndef UNIT_BETA

329

SCALE_BLOCK(1, DATA_TYPE, bias, BETA);

330

#endif // UNIT_BIAS

331

332

// c = c + bias[broadcasted]

333

ADD_BLOCK_BROADCAST(M0, c, bias0);

334

335

#else // defined(BROADCAST_BIAS)

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

336

__global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

337

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

338

LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

339

340

#ifndef UNIT_BETA

341

SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);

#endif // UNIT_BIAS

// c = c + bias

ADD_BLOCK(M0, c, bias);

346

347

#endif // defined(BROADCAST_BIAS)

348

#endif // defined(BETA)

349

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

350

const bool cond_y = y == 0;

351

const bool cond_x = ((x + 1) * N0 >= N);

352

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

353

// c = act(c)

354

POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);

355

// c = c + eltwise_operand (mix-precision, broadcast, boundary aware)

SiCongLi

71cbd28

2021-11-03 12:17:06 +0000

[diff] [blame]

356

POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);

SiCongLi

afa1972

2021-10-24 19:12:33 +0100

[diff] [blame]

357

// c = act(c)

358

POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);

359

360

// Store output block

361