Blame - src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp - ml/ComputeLibrary

Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,

58

const GEMMReshapeInfo &gemm_info)

59

{

60

ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);

61

ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);

62

ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);

63

ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");

64

ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");

65

ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose);

66

ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose);

67

ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);

68

ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");

69

ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);

70

ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);

71

ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");

72

73

const int m = gemm_info.m();

74

const int n = gemm_info.n();

75

const int k = gemm_info.k();

76

77

TensorShape tensor_shape0{ input0->tensor_shape() };

78

tensor_shape0.set(0, k);

79

tensor_shape0.set(1, m);

80

81

TensorShape tensor_shape1{ input1->tensor_shape() };

82

tensor_shape1.set(0, n);

83

tensor_shape1.set(1, k);

84

85

const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);

86

const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);

87

88

const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));

89

const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));

90

91

ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);

92

ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);

93

94

if(output->total_size() != 0)

95

{

96

const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));

97

ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);

98

ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);

}

return Status{};

}

std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,

105

const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)

106

{

107

unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];

108

unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];

109

bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);

Window win{};

Window win_out{};

bool window_changed = false;

114

115

// Output tensor auto initialization if not yet initialized

116

auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)).set_data_type(DataType::S32));

117

118

TensorInfo tmp_info(*output);

119

120

if(reinterpret_output_as_3d)

121

{

122

// Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,

123

// the window needs to be constructed on the 2D collapsed version of the tensor

124

TensorShape tmp_shape(output->tensor_shape());

125

tmp_shape.collapse(2U, 1U);

126

tmp_info.set_tensor_shape(tmp_shape);

127

}

128

129

// Configure kernel window

130

num_elems_processed_per_iteration_x = rhs_info.n0;

131

num_elems_processed_per_iteration_y = lhs_info.m0;

132

133

// Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor

134

// The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic

135

const int m = gemm_info.m();

136

const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;

137

138

win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

139

win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

140

141

AccessWindowStatic input0_access(input0, 0, 0,

142

ceil_to_multiple(input0->dimension(0), num_elems_processed_per_iteration_y),

143

input0->dimension(1));

144

AccessWindowStatic input1_access(input1, 0, 0,

145

ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),

146

input1->dimension(1));

147

AccessWindowStatic output_access(output, 0, 0,

148

ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),

149

output->dimension(1) + bottom_pad);

150

151

window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop

152

update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor

153

154

output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));

155

156

// Collapse along the Z direction

157

// This collapse needs to be here in order to tune the Z dimension of LWS

158

Window collapsed = win;

159

const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);

160

collapsed = win.collapse(win, dimension_to_collapse);

161

162

Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};

163

return std::make_pair(err, collapsed);

}

} // namespace

CLGEMMLowpMatrixMultiplyReshapedKernel::CLGEMMLowpMatrixMultiplyReshapedKernel()

168

: _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1)

{

}

void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,

173

const GEMMReshapeInfo &gemm_info)

174

{

175

ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);

176

177

ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info));

_input0 = input0;

_input1 = input1;

_output = output;

_reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);

183

_k = gemm_info.k();

184

185

// Check if we need to slide the matrix B

186

const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();

187

_slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);

188

189

ElementsProcessed num_elements_processed{};

190

191

// Configure kernel window

192

auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);

193

ARM_COMPUTE_ERROR_THROW_ON(win_config.first);

194

ICLKernel::configure_internal(win_config.second);

195

196

// Create build options

197

CLBuildOptions build_opts;

198

build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");

199

build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));

200

build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));

201

build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));

202

build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");

203

build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");

204

build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));

205

build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));

206

build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));

207

build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));

208

build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));

209

210

std::string kernel_name("gemmlowp_mm_reshaped_");

211

kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";

212

kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";

213

kernel_name += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";

214

215

// Create kernel

216

_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));

217

218

// Set config_id for enabling LWS tuning

219

_config_id = kernel_name;

220

_config_id += "_";

221

_config_id += (_reinterpret_output_as_3d ? "3do_" : "");

222

_config_id += support::cpp11::to_string(output->info()->dimension(1));

223

_config_id += "_";

224

_config_id += support::cpp11::to_string(output->info()->dimension(0));

225

_config_id += "_";

226

_config_id += support::cpp11::to_string(gemm_info.k());

227

_config_id += "_";

228

_config_id += support::cpp11::to_string(output->info()->dimension(2));

229

_config_id += "_";

230

_config_id += support::cpp11::to_string(lhs_info.m0);

231

_config_id += "_";

232

_config_id += support::cpp11::to_string(rhs_info.n0);

233

_config_id += "_";

234

_config_id += support::cpp11::to_string(lhs_info.k0);

235

_config_id += "_";

236

_config_id += support::cpp11::to_string(lhs_info.v0);

237

_config_id += "_";

238

_config_id += support::cpp11::to_string(rhs_info.h0);

239

_config_id += "_";

240

_config_id += support::cpp11::to_string(lhs_info.interleave);

241

_config_id += "_";

242

_config_id += support::cpp11::to_string(rhs_info.interleave);

243

}

244

245

Status CLGEMMLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,

246

const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)

247

{

248

ElementsProcessed num_elements_processed{};

249

ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, lhs_info, rhs_info, gemm_info));

250

ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),

251

input1->clone().get(),

252

output->clone().get(),

lhs_info,

rhs_info,

gemm_info,

num_elements_processed)

.first);

return Status{};

}

void CLGEMMLowpMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQueue &queue)

263

{

264

ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);

265

ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

266

267

if(_input1->info()->num_dimensions() < 3)

268

{

269

// The stride_z for matrix B must be zero if we do not slice

270

ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);

271

}

272

273

Window slice = window.first_slice_window_3D();

274

Window slice_matrix_b = slice;

275

276

slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));

277

slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));

278

279

if(_reinterpret_output_as_3d)

280

{

281

// Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor

282

const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 4;

283

const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;

284

_kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));

}

do

{

Window slice_b = slice;

290

// Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2

291

// This scenario can happen when the matrix multiplication is used to perform a convolution operation

292

if(!_slide_matrix_b)

293

{

294

slice_b = slice_matrix_b;

295

}

296

297

unsigned int idx = 0;

298

add_2D_tensor_argument(idx, _input0, slice);

299

add_2D_tensor_argument(idx, _input1, slice_b);

300

add_2D_tensor_argument(idx, _output, slice);

301

_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));

302

_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));

303

_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));

304

_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));

305

enqueue(queue, *this, slice, lws_hint());

306

}

307

while(window.slide_window_slice_3D(slice));

308

}