GeezerVison is showing github.com — Fuser/python/python_common/python_utils.cpp at main · NVIDIA/Fuser

Expand file tree

python_utils.cpp

Copy path

More file actions

Latest commit

History

294 lines (273 loc) · 9.15 KB

python_utils.cpp

Copy path

File metadata and controls

294 lines (273 loc) · 9.15 KB

Raw

Copy raw file

Download raw file

Open symbols panel

Edit and raw actions

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

// clang-format off

* SPDX-License-Identifier: BSD-3-Clause

// clang-format on

#include <python_utils.h>

#include <polymorphic_value.h>

#include <algorithm>

#include <ranges>

namespace nvfuser {

struct DimInfo {

int64_t index;

int64_t size;

int64_t stride;

int64_t stride_order;

std::optional<bool> contiguity = std::nullopt;

bool isBroadcast() {

return stride == 0 || size == 1;

}

};

std::pair<std::vector<std::optional<bool>>, std::vector<int64_t>>

computeTensorDescriptor(

const std::vector<int64_t>& sizes,

const std::vector<int64_t>& strides) {

NVF_CHECK(

sizes.size() == strides.size(),

"compute_tensor_descriptor: "

"Sizes and strides must have the same number of dimensions");

std::vector<DimInfo> non_broadcast_dim_info_vec;

std::vector<DimInfo> stride_zero_dims;

for (auto i : std::ranges::iota_view(0u, sizes.size())) {

// NOTE: not supporting negative stride yet, but we can probably allow it on

// broadcast dims

NVF_CHECK(

strides[i] >= 0,

"negative stride on tensor is not supported: strides[",

"]=",

strides[i]);

DimInfo dim_info{(int64_t)i, sizes[i], strides[i]};

if (strides[i] != 0) {

non_broadcast_dim_info_vec.push_back(dim_info);

} else {

stride_zero_dims.push_back(dim_info);

}

// sort non-broadcast dimensions by stride

std::stable_sort(

non_broadcast_dim_info_vec.begin(),

non_broadcast_dim_info_vec.end(),

[](const auto& l, const auto& r) { return l.stride > r.stride; });

// combine dimensions while preserving the semantical position of broadcast

// dimensions

for (const auto& dim_info : stride_zero_dims) {

non_broadcast_dim_info_vec.insert(

non_broadcast_dim_info_vec.begin() + dim_info.index, dim_info);

}

// Dimensions are marked contiguous by inspecting the current dimension and

// one to the right towards the inner dimension while skipping over broadcast

// dimensions.

// The innermost dimension, that is not broadcasted, does not have any

// dimension to it's right and needs to have stride equal to 1 in order to be

// marked contiguous.

for (int64_t i = 0; i < (int64_t)sizes.size();) {

non_broadcast_dim_info_vec[i].stride_order = (int64_t)sizes.size() - 1 - i;

if (!non_broadcast_dim_info_vec[i].isBroadcast()) {

auto l = i++;

int64_t expected = 1;

for (; i < (int64_t)sizes.size(); i++) {

non_broadcast_dim_info_vec[i].stride_order =

(int64_t)sizes.size() - 1 - i;

if (!non_broadcast_dim_info_vec[i].isBroadcast()) {

expected = non_broadcast_dim_info_vec[i].stride *

non_broadcast_dim_info_vec[i].size;

break;

}

non_broadcast_dim_info_vec[l].contiguity =

(non_broadcast_dim_info_vec[l].stride == expected);

} else {

i++;

}

std::vector<int64_t> stride_order_vec(sizes.size(), -1);

for (const auto& dim_info : non_broadcast_dim_info_vec) {

stride_order_vec[dim_info.index] = dim_info.stride_order;

}

std::vector<std::optional<bool>> contiguity_vec;

std::transform(

non_broadcast_dim_info_vec.begin(),

non_broadcast_dim_info_vec.end(),

std::back_inserter(contiguity_vec),

[](const DimInfo& val) { return val.contiguity; });

return std::make_pair(contiguity_vec, stride_order_vec);

}

void verifyShape(const std::vector<int64_t>& shape) {

for (size_t i = 0; i < shape.size(); ++i) {

NVF_CHECK(

shape[i] >= -1,

"The value ",

shape[i],

" at index ",

" was neither symbolic(-1), zero_element(0), broadcast(1), or "

"static(>1).");

}

void normalizeStrideOrder(std::vector<int64_t>& stride_order) {

if (stride_order.empty()) {

return;

}

int64_t rank = (int64_t)stride_order.size();

// Validate first

std::for_each(

stride_order.begin(), stride_order.end(), [rank](int64_t order) {

if (order < 0) {

NVF_CHECK(

order >= -rank,

"defineTensor stride_order argument is out of range, expects >= ",

-rank,

", but got: ",

order);

} else {

NVF_CHECK(

order < rank,

"defineTensor stride_order argument is out of range, expects < ",

rank,

", but got: ",

order);

}

});

// Then, normalize negative values.

std::unordered_set<int64_t> order_set;

order_set.reserve(rank);

std::transform(

stride_order.begin(),

stride_order.end(),

std::inserter(order_set, order_set.begin()),

[rank](int64_t order) { return wrapDim(order, rank); });

NVF_CHECK(

order_set.size() == stride_order.size(),

"defineTensor got duplicated stride_order entries: " +

toDelimitedString(stride_order));

}

std::vector<bool> getExpanded(

const std::vector<int64_t>& shape,

const std::vector<std::optional<bool>>& contiguity,

const std::vector<int64_t>& stride_order) {

NVF_CHECK(

contiguity.size() == shape.size(),

"Length of contiguity argument (",

contiguity.size(),

") must match that of shape argument (",

shape.size(),

")");

NVF_CHECK(

stride_order.empty() || stride_order.size() == shape.size(),

"Length of stride_order argument (",

stride_order.size(),

") must be zero or match that of shape argument (",

shape.size(),

")");

size_t rank = shape.size();

std::vector<bool> is_expand(rank);

for (size_t index : arange(rank)) {

// since contiguity vector is given to the corresponding order in alloc

// domain, while is_expand is given to root domain, we need to map it

// correctly with `contig_index` and `index`.

// stride_order[i] indicates that:

// `logical_domain[i]` maps to `alloc_domain[rank - 1 - stride_order_[i]]`

// Hence `index` on root domain would be corresponding to the contiguity

// index `contig_index = rank - 1 - stride_order[index]`

const size_t contig_index = stride_order.empty()

? index

: rank - 1 - static_cast<size_t>(stride_order.at(index));

const bool is_broadcast = !contiguity.at(contig_index).has_value();

const bool has_non_broadcast_size = (shape.at(index) != 1);

// A root dimension is expand dimension if:

// The dimension is marked a broadcast; and

// The dimension has an expanded extent.

is_expand[index] = is_broadcast && has_non_broadcast_size;

}

return is_expand;

}

std::vector<std::optional<bool>> getContiguityVec(

const std::vector<int64_t>& shape,

const std::vector<int64_t>& stride_order,

bool contiguity) {

const auto rank = static_cast<int64_t>(shape.size());

std::vector<std::optional<bool>> contiguity_vec(rank);

for (const auto index : arange(rank)) {

const auto contig_index =

stride_order.empty() ? index : rank - 1 - stride_order[index];

if (shape[index] == 1) {

contiguity_vec[contig_index] = std::nullopt;

} else {

contiguity_vec[contig_index] = contiguity;

}

return contiguity_vec;

}

std::vector<int64_t> getTensorViewBuilderSizes(

const std::vector<int64_t>& sizes,

bool static_sizes) {

// TensorViewBuilder assumes any dim with a compile-time constant size == 1

// is a broadcast axis and symbolic sizes are identified by -1.

std::vector<int64_t> dim_sizes;

dim_sizes.reserve(sizes.size());

for (size_t i : arange(sizes.size())) {

NVF_ERROR(

sizes[i] >= 0,

"Size of ",

sizes[i],

" is not supported in nvFuser. Expected size >= 0.");

if (static_sizes) {

dim_sizes.push_back(sizes[i]);

} else { // Symbolic defined tensor for dynamic shape usage

if (sizes[i] == 1) {

dim_sizes.push_back(1);

} else {

dim_sizes.push_back(-1);

}

return dim_sizes;

}

const char* dtypeToPyString(PrimDataType t) {

switch (t) {

case DataType::Bool:

return "DataType.Bool";

case DataType::Double:

return "DataType.Double";

case DataType::Float:

return "DataType.Float";

case DataType::Half:

return "DataType.Half";

case DataType::BFloat16:

return "DataType.BFloat16";

case DataType::Float8_e4m3fn:

return "DataType.Float8_e4m3fn";

case DataType::Float8_e5m2:

return "DataType.Float8_e5m2";

case DataType::Float8_e8m0fnu:

return "DataType.Float8_e8m0fnu";

case DataType::Float4_e2m1fn:

return "DataType.Float4_e2m1fn";

case DataType::Float4_e2m1fn_x2:

return "DataType.Float4_e2m1fn_x2";

case DataType::Index:

return "DataType.Index";

case DataType::Int:

return "DataType.Int";

case DataType::Int32:

return "DataType.Int32";

case DataType::ComplexFloat:

return "DataType.ComplexFloat";

case DataType::ComplexDouble:

return "DataType.ComplexDouble";

case DataType::Null:

return "DataType.Null";

case DataType::UInt64:

return "DataType.UInt64";

default:

break;

}

NVF_THROW("No string found for data type.");

return nullptr;

}

} // namespace nvfuser