| 1 | //===- NVGPUDialect.cpp - MLIR NVGPU ops implementation -------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements the NVGPU dialect and its operations. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/NVGPU/IR/NVGPUDialect.h" |
| 14 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 15 | #include "mlir/Dialect/MemRef/IR/MemoryAccessOpInterfaces.h" |
| 16 | #include "mlir/IR/Builders.h" |
| 17 | #include "mlir/IR/BuiltinAttributes.h" |
| 18 | #include "mlir/IR/BuiltinTypes.h" |
| 19 | #include "mlir/IR/Diagnostics.h" |
| 20 | #include "mlir/IR/TypeUtilities.h" |
| 21 | #include "mlir/IR/Verifier.h" |
| 22 | #include "llvm/ADT/STLExtras.h" |
| 23 | |
| 24 | using namespace mlir; |
| 25 | using namespace mlir::nvgpu; |
| 26 | |
| 27 | #include "mlir/Dialect/NVGPU/IR/NVGPUDialect.cpp.inc" |
| 28 | |
| 29 | void NVGPUDialect::initialize() { |
| 30 | addTypes< |
| 31 | #define GET_TYPEDEF_LIST |
| 32 | #include "mlir/Dialect/NVGPU/IR/NVGPUTypeDefs.cpp.inc" |
| 33 | >(); |
| 34 | addAttributes< |
| 35 | #define GET_ATTRDEF_LIST |
| 36 | #include "mlir/Dialect/NVGPU/IR/NVGPUAttrDefs.cpp.inc" |
| 37 | >(); |
| 38 | addOperations< |
| 39 | #define GET_OP_LIST |
| 40 | #include "mlir/Dialect/NVGPU/IR/NVGPUOps.cpp.inc" |
| 41 | >(); |
| 42 | declarePromisedInterfaces<memref::IndexedAccessOpInterface, LdMatrixOp>(); |
| 43 | declarePromisedInterfaces<memref::IndexedMemCopyOpInterface, |
| 44 | DeviceAsyncCopyOp>(); |
| 45 | } |
| 46 | |
| 47 | bool NVGPUDialect::isSharedMemoryAddressSpace(Attribute memorySpace) { |
| 48 | if (!memorySpace) |
| 49 | return false; |
| 50 | if (auto intAttr = llvm::dyn_cast<IntegerAttr>(Val&: memorySpace)) |
| 51 | return intAttr.getValue().getZExtValue() == |
| 52 | NVGPUDialect::kSharedMemoryAddressSpace; |
| 53 | if (auto gpuAttr = llvm::dyn_cast<gpu::AddressSpaceAttr>(Val&: memorySpace)) |
| 54 | return gpuAttr.getValue() == gpu::AddressSpace::Workgroup; |
| 55 | return false; |
| 56 | } |
| 57 | |
| 58 | bool NVGPUDialect::hasSharedMemoryAddressSpace(MemRefType type) { |
| 59 | Attribute memorySpace = type.getMemorySpace(); |
| 60 | return isSharedMemoryAddressSpace(memorySpace); |
| 61 | } |
| 62 | |
| 63 | //===----------------------------------------------------------------------===// |
| 64 | // NVGPU_DeviceAsyncCopyOp |
| 65 | //===----------------------------------------------------------------------===// |
| 66 | |
| 67 | LogicalResult DeviceAsyncCopyOp::verify() { |
| 68 | auto srcMemref = llvm::cast<MemRefType>(Val: getSrc().getType()); |
| 69 | auto dstMemref = llvm::cast<MemRefType>(Val: getDst().getType()); |
| 70 | |
| 71 | if (!srcMemref.isLastDimUnitStride()) |
| 72 | return emitError(message: "source memref most minor dim must have unit stride" ); |
| 73 | if (!dstMemref.isLastDimUnitStride()) |
| 74 | return emitError(message: "destination memref most minor dim must have unit stride" ); |
| 75 | if (!NVGPUDialect::hasSharedMemoryAddressSpace(type: dstMemref)) |
| 76 | return emitError() |
| 77 | << "destination memref must have a memory space attribute of " |
| 78 | "IntegerAttr(" |
| 79 | << NVGPUDialect::kSharedMemoryAddressSpace |
| 80 | << ") or gpu::AddressSpaceAttr(Workgroup)" ; |
| 81 | if (dstMemref.getElementType() != srcMemref.getElementType()) |
| 82 | return emitError(message: "source and destination must have the same element type" ); |
| 83 | if (size_t(srcMemref.getRank()) != getSrcIndices().size()) |
| 84 | return emitOpError() << "expected " << srcMemref.getRank() |
| 85 | << " source indices, got " << getSrcIndices().size(); |
| 86 | if (size_t(dstMemref.getRank()) != getDstIndices().size()) |
| 87 | return emitOpError() << "expected " << dstMemref.getRank() |
| 88 | << " destination indices, got " |
| 89 | << getDstIndices().size(); |
| 90 | int64_t dstElements = getDstElements().getZExtValue(); |
| 91 | int64_t sizeInBytes = (dstMemref.getElementTypeBitWidth() * dstElements) / 8; |
| 92 | if (sizeInBytes != 4 && sizeInBytes != 8 && sizeInBytes != 16) { |
| 93 | unsigned dstWidth = dstMemref.getElementTypeBitWidth(); |
| 94 | InFlightDiagnostic diag = emitError(); |
| 95 | diag << "Requested copy elements is " << dstElements << " with width " |
| 96 | << dstMemref.getElementTypeBitWidth() |
| 97 | << ". But copy elements could be one of " ; |
| 98 | if ((32 / dstWidth) > 0) |
| 99 | diag << (32 / dstWidth) << ", " ; |
| 100 | if ((64 / dstWidth) > 0) |
| 101 | diag << (64 / dstWidth) << ", " ; |
| 102 | if ((128 / dstWidth) > 0) |
| 103 | diag << (128 / dstWidth) << "." ; |
| 104 | return diag; |
| 105 | } |
| 106 | if (getBypassL1().has_value()) { |
| 107 | int64_t req = 16 * 8 / dstMemref.getElementTypeBitWidth(); |
| 108 | if (getBypassL1().value() && sizeInBytes != 16) { |
| 109 | return emitOpError() << "bypassL1 does not satify alignment for " |
| 110 | << dstMemref << " with destination element " |
| 111 | << dstElements |
| 112 | << ". Unset bypassL1, or set " |
| 113 | "destination element to " |
| 114 | << req; |
| 115 | } |
| 116 | } |
| 117 | return success(); |
| 118 | } |
| 119 | |
| 120 | //===----------------------------------------------------------------------===// |
| 121 | // NVGPU_MmaSyncOp |
| 122 | //===----------------------------------------------------------------------===// |
| 123 | void MmaSyncOp::build(::mlir::OpBuilder &odsBuilder, |
| 124 | ::mlir::OperationState &odsState, Value matrixA, |
| 125 | Value matrixB, Value matrixC, ArrayAttr mmaShape) { |
| 126 | build(odsBuilder, odsState, res: matrixC.getType(), matrixA, matrixB, matrixC, |
| 127 | mmaShape, tf32Enabled: UnitAttr()); |
| 128 | } |
| 129 | |
| 130 | void MmaSyncOp::build(::mlir::OpBuilder &odsBuilder, |
| 131 | ::mlir::OperationState &odsState, Value matrixA, |
| 132 | Value matrixB, Value matrixC, ArrayRef<int64_t> mmaShape, |
| 133 | bool tf32Enabled) { |
| 134 | build(odsBuilder, odsState, res: matrixC.getType(), matrixA, matrixB, matrixC, |
| 135 | mmaShape: odsBuilder.getI64ArrayAttr(values: mmaShape), |
| 136 | tf32Enabled: tf32Enabled ? odsBuilder.getUnitAttr() : UnitAttr()); |
| 137 | } |
| 138 | |
| 139 | /// Performs verification for MmaSyncOp and MmaSparseSyncOp. |
| 140 | static LogicalResult verifyMmaSyncOp(Operation *op, |
| 141 | TypedValue<VectorType> matrixA, |
| 142 | TypedValue<VectorType> matrixB, |
| 143 | TypedValue<VectorType> matrixC, |
| 144 | const std::array<int64_t, 3> &mmaShape, |
| 145 | bool tf32Enabled, bool sparse = false) { |
| 146 | // The verification for mma.sync covering various shapes and data types is |
| 147 | // based on the fundamental tensor core shape. |
| 148 | |
| 149 | // "Fundamental" tensor core shapes: |
| 150 | // - For F32 (TF32), F16, S8, and S4 data |
| 151 | // types the fundamental tensor core operation is of shape 8-by-8-by-128b. |
| 152 | // - F64 is an exception and is of shape 8-by-8-by-256b. |
| 153 | int64_t shapeM = 8; |
| 154 | int64_t shapeN = 8; |
| 155 | int64_t shapeK; // set based on data type (128b for all data types except F64) |
| 156 | |
| 157 | // Number of elements A, B, and C per thread per fundamental tensor core tile |
| 158 | int64_t numElementA; // set based on data type (32b except F64) |
| 159 | int64_t numElementB; // set based on data type (32b except F64) |
| 160 | int64_t numElementC{2}; // two accumulator elements per fundamental tile |
| 161 | |
| 162 | // nvgpu.mma.sync vector operands (per thread) |
| 163 | auto aVector = matrixA.getType(); |
| 164 | auto bVector = matrixB.getType(); |
| 165 | auto cVector = matrixC.getType(); |
| 166 | |
| 167 | // vector shapes |
| 168 | ArrayRef<int64_t> aShape = aVector.getShape(); |
| 169 | ArrayRef<int64_t> bShape = bVector.getShape(); |
| 170 | ArrayRef<int64_t> cShape = cVector.getShape(); |
| 171 | |
| 172 | // vector element type |
| 173 | Type aType = aVector.getElementType(); |
| 174 | |
| 175 | // Certain data types are not allowed in sparse mode. |
| 176 | if (sparse && aType.isF64()) |
| 177 | return op->emitError() << "f64 is not supported for sparse mode" ; |
| 178 | |
| 179 | if (aType.isF64()) { |
| 180 | // exception to 8-by-8-128b fundamental tensor core tile size |
| 181 | shapeK = 4; |
| 182 | numElementA = 1; |
| 183 | numElementB = 1; |
| 184 | } else if (aType.isF32() || aType.isBF16() || aType.isF16() || |
| 185 | aType.isInteger(width: 8) || aType.isInteger(width: 4) || aType.isF8E4M3FN() || |
| 186 | aType.isF8E5M2()) { |
| 187 | // 8-by-8-128b fundamental tensor core tile size |
| 188 | int operandBitwidth = aType.getIntOrFloatBitWidth(); |
| 189 | shapeK = 128 / operandBitwidth; // 128b wide shapeK |
| 190 | |
| 191 | numElementA = 32 / operandBitwidth; // 32b wide operand A |
| 192 | numElementB = 32 / operandBitwidth; // 32b wide operand B |
| 193 | } else { |
| 194 | return op->emitError() |
| 195 | << "expected input data type (i4,i8,f16,bf16,tf32,f64," |
| 196 | "f8E4M3FN,f8E5M2) supported by " |
| 197 | << op->getName(); |
| 198 | } |
| 199 | |
| 200 | // |
| 201 | // Basic verification |
| 202 | // |
| 203 | |
| 204 | if (aShape.size() != 2) { |
| 205 | return op->emitError() << "matrixA must be 2 dimensional vector" ; |
| 206 | } |
| 207 | |
| 208 | if (bShape.size() != 2) { |
| 209 | return op->emitError() << "matrixB must be 2 dimensional vector" ; |
| 210 | } |
| 211 | |
| 212 | if (cShape.size() != 2) { |
| 213 | return op->emitError() << "matrixC must be 2 dimensional vector" ; |
| 214 | } |
| 215 | |
| 216 | auto [m, n, k] = mmaShape; |
| 217 | |
| 218 | // verify warp-wide size for vector a |
| 219 | int64_t sparseFactor = sparse ? 2 : 1; |
| 220 | if (aShape[0] * aShape[1] * kWarpSize != m * k / sparseFactor) |
| 221 | return op->emitOpError() |
| 222 | << "expected " << m * k << " warp-wide matrix A elements" ; |
| 223 | |
| 224 | // verify warp-wide size for vector b |
| 225 | if (bShape[0] * bShape[1] * kWarpSize != k * n) |
| 226 | return op->emitOpError() |
| 227 | << "expected " << k * n << " warp-wide matrix B elements" ; |
| 228 | |
| 229 | // verify warp-wide size for vector c |
| 230 | if (cShape[0] * cShape[1] * kWarpSize != m * n) |
| 231 | return op->emitOpError() |
| 232 | << "expected " << m * n << " warp-wide matrix C elements" ; |
| 233 | |
| 234 | // verify tf32 tensor cores are enabled for only F32 datatype |
| 235 | if (tf32Enabled && !(aType.isF32())) |
| 236 | return op->emitOpError() |
| 237 | << "expected tf32 tensor cores only for F32 operands" ; |
| 238 | |
| 239 | // |
| 240 | // Extended verification |
| 241 | // |
| 242 | |
| 243 | // tiles of fundamental tensor core operations |
| 244 | int64_t mTile = m / shapeM; |
| 245 | int64_t nTile = n / shapeN; |
| 246 | int64_t kTile = k / shapeK; |
| 247 | |
| 248 | // verify shape of aVector |
| 249 | if ((aShape[0] != mTile * kTile / (sparse ? 2 : 1)) || |
| 250 | (aShape[1] != numElementA)) |
| 251 | return op->emitOpError() << "expected matrix A to be shaped (" |
| 252 | << mTile * kTile << " x " << numElementA << ")" ; |
| 253 | |
| 254 | // verify shape of bVector |
| 255 | if ((bShape[0] != kTile * nTile) || (bShape[1] != numElementB)) |
| 256 | return op->emitOpError() << "expected matrix B to be shaped (" |
| 257 | << kTile * nTile << " x " << numElementB << ")" ; |
| 258 | |
| 259 | // verify shape of cVector |
| 260 | if ((cShape[0] != mTile * nTile) || (cShape[1] != numElementC)) |
| 261 | return op->emitOpError() << "expected matrix C to be shaped (" |
| 262 | << mTile * nTile << " x " << numElementC << ")" ; |
| 263 | |
| 264 | return success(); |
| 265 | } |
| 266 | |
| 267 | LogicalResult MmaSyncOp::verify() { |
| 268 | if (getMmaShape().size() != 3) |
| 269 | return emitOpError() << "mmaShape must have exactly 3 elements" ; |
| 270 | |
| 271 | return verifyMmaSyncOp(op: this->getOperation(), matrixA: getMatrixA(), matrixB: getMatrixB(), |
| 272 | matrixC: getMatrixC(), mmaShape: getMmaShapeAsArray(), |
| 273 | tf32Enabled: getTf32Enabled().value_or(u: false)); |
| 274 | } |
| 275 | |
| 276 | //===----------------------------------------------------------------------===// |
| 277 | // NVGPU_MmaSparseSyncOp |
| 278 | //===----------------------------------------------------------------------===// |
| 279 | void MmaSparseSyncOp::build(::mlir::OpBuilder &odsBuilder, |
| 280 | ::mlir::OperationState &odsState, Value matrixA, |
| 281 | Value matrixB, Value matrixC, Value sparseMetadata, |
| 282 | ArrayRef<int64_t> mmaShape) { |
| 283 | build(odsBuilder, odsState, res: matrixC.getType(), matrixA, matrixB, matrixC, |
| 284 | sparseMetadata, mmaShape: odsBuilder.getI64ArrayAttr(values: mmaShape), sparsitySelector: 0, tf32Enabled: UnitAttr()); |
| 285 | } |
| 286 | |
| 287 | LogicalResult MmaSparseSyncOp::verify() { |
| 288 | unsigned sparsitySelector = getSparsitySelector(); |
| 289 | if (sparsitySelector > 1) |
| 290 | return emitOpError() << "sparsity selector should be 0 or 1" ; |
| 291 | |
| 292 | if (getMmaShape().size() != 3) |
| 293 | return emitOpError() << "mmaShape must have exactly 3 elements" ; |
| 294 | |
| 295 | return verifyMmaSyncOp(op: this->getOperation(), matrixA: getMatrixA(), matrixB: getMatrixB(), |
| 296 | matrixC: getMatrixC(), mmaShape: getMmaShapeAsArray(), |
| 297 | tf32Enabled: getTf32Enabled().value_or(u: false), sparse: true); |
| 298 | } |
| 299 | |
| 300 | //===----------------------------------------------------------------------===// |
| 301 | // NVGPU_LdMatrixOp |
| 302 | //===----------------------------------------------------------------------===// |
| 303 | LogicalResult LdMatrixOp::verify() { |
| 304 | // ldmatrix reads data from source in shared memory |
| 305 | auto srcMemref = llvm::cast<MemRefType>(Val: getSrcMemref().getType()); |
| 306 | |
| 307 | // ldmatrix writes data to result/destination in vector registers |
| 308 | auto resVector = llvm::cast<VectorType>(Val: getRes().getType()); |
| 309 | |
| 310 | // vector register shape, element type, and bitwidth |
| 311 | ArrayRef<int64_t> resShape = resVector.getShape(); |
| 312 | Type resType = resVector.getElementType(); |
| 313 | int64_t elementBitWidth = resType.getIntOrFloatBitWidth(); |
| 314 | |
| 315 | // ldmatrix loads 32 bits into vector registers per 8-by-8 tile per thread |
| 316 | int64_t numElementsPer32b = 32 / elementBitWidth; |
| 317 | |
| 318 | // number of 8-by-8 tiles |
| 319 | int64_t numTiles = getNumTiles(); |
| 320 | |
| 321 | // transpose elements in vector registers at 16b granularity when true |
| 322 | bool isTranspose = getTranspose(); |
| 323 | |
| 324 | // |
| 325 | // verification |
| 326 | // |
| 327 | |
| 328 | if (!NVGPUDialect::hasSharedMemoryAddressSpace(type: srcMemref)) |
| 329 | return emitError() |
| 330 | << "expected nvgpu.ldmatrix srcMemref must have a memory space " |
| 331 | "attribute of IntegerAttr(" |
| 332 | << NVGPUDialect::kSharedMemoryAddressSpace |
| 333 | << ") or gpu::AddressSpaceAttr(Workgroup)" ; |
| 334 | if (elementBitWidth > 32) |
| 335 | return emitError() << "nvgpu.ldmatrix works for 32b or lower" ; |
| 336 | if (isTranspose && !(elementBitWidth == 16)) |
| 337 | return emitError() |
| 338 | << "nvgpu.ldmatrix transpose works only at 16b granularity" ; |
| 339 | if (resShape.size() != 2) { |
| 340 | return emitError() << "results must be 2 dimensional vector" ; |
| 341 | } |
| 342 | if (!(resShape[1] == numElementsPer32b)) |
| 343 | return emitError() << "expected vector register shape[1] = " |
| 344 | << numElementsPer32b; |
| 345 | if (!(resShape[0] == numTiles)) |
| 346 | return emitError() |
| 347 | << "expected vector register shape[0] and numTiles to match" ; |
| 348 | |
| 349 | return success(); |
| 350 | } |
| 351 | |
| 352 | //===----------------------------------------------------------------------===// |
| 353 | // NVGPU_TmaAsyncLoadOp |
| 354 | //===----------------------------------------------------------------------===// |
| 355 | |
| 356 | static unsigned getSwizzleBytes(TensorMapSwizzleKind kind) { |
| 357 | switch (kind) { |
| 358 | case TensorMapSwizzleKind::SWIZZLE_32B: |
| 359 | return 32; |
| 360 | case TensorMapSwizzleKind::SWIZZLE_64B: |
| 361 | return 64; |
| 362 | case TensorMapSwizzleKind::SWIZZLE_128B: |
| 363 | return 128; |
| 364 | default: |
| 365 | return 0; |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | std::optional<InFlightDiagnostic> verifyTmaDescriptorWithMemref( |
| 370 | Operation *op, TensorMapDescriptorType descType, |
| 371 | std::optional<MemRefType> memrefType = std::nullopt) { |
| 372 | MemRefType descMemref = descType.getTensor(); |
| 373 | // Limitation |
| 374 | if (descType.getInterleave() != TensorMapInterleaveKind::INTERLEAVE_NONE) |
| 375 | return op->emitError() << "Interleave options are not supported yet." ; |
| 376 | |
| 377 | // Address space check for shared memory check |
| 378 | if (!NVGPUDialect::hasSharedMemoryAddressSpace(type: descMemref)) { |
| 379 | return op->emitError() << "the tensor map descriptor has incorrect address " |
| 380 | "space, it must be shared memory address space." ; |
| 381 | } |
| 382 | // Support only static shape for the time being |
| 383 | if (!descMemref.hasStaticShape()) |
| 384 | return op->emitError() << "the tensor map descriptor must be static shaped" ; |
| 385 | |
| 386 | for (auto dim : descMemref.getShape()) { |
| 387 | if (dim <= 0 || dim > kMaxTMADimension) { |
| 388 | return op->emitError() << "the tensor map descriptor must have " |
| 389 | "dimensions between 1 and " |
| 390 | << kMaxTMADimension << " but it is " << dim; |
| 391 | } |
| 392 | } |
| 393 | if (descMemref.getRank() > 1 && |
| 394 | descType.getSwizzle() != TensorMapSwizzleKind::SWIZZLE_NONE) { |
| 395 | unsigned lastDimensionByte = |
| 396 | descMemref.getElementTypeBitWidth() * descMemref.getShape().back() / 8; |
| 397 | unsigned expectByte = getSwizzleBytes(kind: descType.getSwizzle()); |
| 398 | if (lastDimensionByte != expectByte) |
| 399 | return op->emitError() << "the tensormap descriptor must have last " |
| 400 | "dimension of " |
| 401 | << expectByte << " bytes but it is " |
| 402 | << lastDimensionByte << " bytes" ; |
| 403 | } |
| 404 | |
| 405 | // No verification if memref type is not provided |
| 406 | if (!memrefType.has_value()) |
| 407 | return std::nullopt; |
| 408 | |
| 409 | MemRefType dstMemref = memrefType.value(); |
| 410 | |
| 411 | // Check element type |
| 412 | if (descMemref.getElementType() != dstMemref.getElementType()) { |
| 413 | return op->emitError() << "the element type of tensor map descriptor and " |
| 414 | "memref must be same" ; |
| 415 | } |
| 416 | |
| 417 | if (!NVGPUDialect::hasSharedMemoryAddressSpace(type: dstMemref)) { |
| 418 | return op->emitError() << "the destination memref has incorrect address " |
| 419 | "space, it must be shared memory address space." ; |
| 420 | } |
| 421 | if (!dstMemref.hasStaticShape()) |
| 422 | return op->emitError() << "the destination memref must be static shaped" ; |
| 423 | |
| 424 | if (dstMemref.getRank() != descMemref.getRank()) { |
| 425 | return op->emitError() << "the shape of tensor map descriptor and " |
| 426 | "memref must have same rank" ; |
| 427 | } |
| 428 | if (!descMemref.getShape().equals(RHS: dstMemref.getShape())) { |
| 429 | return op->emitError() << "memref and tensor map shapes mismatch " |
| 430 | << descMemref << " != " << dstMemref; |
| 431 | } |
| 432 | |
| 433 | int lastDimBytes = |
| 434 | descMemref.getShape().back() * descMemref.getElementTypeBitWidth() / 8; |
| 435 | if (lastDimBytes % kTMALastdimByte != 0) { |
| 436 | return op->emitError() << "the bytes in the last dimension of the tensor " |
| 437 | "map must be a multiple of 16" ; |
| 438 | } |
| 439 | return std::nullopt; |
| 440 | } |
| 441 | |
| 442 | LogicalResult TmaAsyncLoadOp::verify() { |
| 443 | std::optional<InFlightDiagnostic> error = verifyTmaDescriptorWithMemref( |
| 444 | op: *this, descType: getTensorMapDescriptor().getType(), memrefType: getDst().getType()); |
| 445 | if (error.has_value()) |
| 446 | return error.value(); |
| 447 | |
| 448 | if (getCoordinates().size() > kMaxTMATensorDimension) { |
| 449 | return emitError() << "Maximum " << kMaxTMATensorDimension |
| 450 | << " coordinates are supported." ; |
| 451 | } |
| 452 | if (getCoordinates().size() != |
| 453 | size_t(getTensorMapDescriptor().getType().getTensor().getRank())) { |
| 454 | return emitError() << "number of coordinates do not match with the rank of " |
| 455 | "tensor descriptor map." ; |
| 456 | } |
| 457 | |
| 458 | return success(); |
| 459 | } |
| 460 | |
| 461 | //===----------------------------------------------------------------------===// |
| 462 | // NVGPU_TmaAsyncStoreOp |
| 463 | //===----------------------------------------------------------------------===// |
| 464 | |
| 465 | LogicalResult TmaAsyncStoreOp::verify() { |
| 466 | std::optional<InFlightDiagnostic> error = verifyTmaDescriptorWithMemref( |
| 467 | op: *this, descType: getTensorMapDescriptor().getType(), memrefType: getSrc().getType()); |
| 468 | if (error.has_value()) |
| 469 | return error.value(); |
| 470 | |
| 471 | if (getCoordinates().size() > kMaxTMATensorDimension) { |
| 472 | return emitError() << "Maximum " << kMaxTMATensorDimension |
| 473 | << " coordinates are supported." ; |
| 474 | } |
| 475 | if (getCoordinates().size() != |
| 476 | size_t(getTensorMapDescriptor().getType().getTensor().getRank())) { |
| 477 | return emitError() << "number of coordinates do not match with the rank of " |
| 478 | "tensor descriptor map." ; |
| 479 | } |
| 480 | |
| 481 | return success(); |
| 482 | } |
| 483 | |
| 484 | LogicalResult TmaCreateDescriptorOp::verify() { |
| 485 | if (getBoxDimensions().size() > kMaxTMATensorDimension) { |
| 486 | return emitError() << "Maximum " << kMaxTMATensorDimension |
| 487 | << " coordinates are supported." ; |
| 488 | } |
| 489 | |
| 490 | std::optional<InFlightDiagnostic> error = |
| 491 | verifyTmaDescriptorWithMemref(op: *this, descType: getTensorMap().getType()); |
| 492 | if (error.has_value()) |
| 493 | return error.value(); |
| 494 | |
| 495 | return success(); |
| 496 | } |
| 497 | |
| 498 | //===----------------------------------------------------------------------===// |
| 499 | // NVGPU_WarpgroupGenerateDescriptorOp |
| 500 | //===----------------------------------------------------------------------===// |
| 501 | |
| 502 | LogicalResult WarpgroupGenerateDescriptorOp::verify() { |
| 503 | std::optional<InFlightDiagnostic> error = |
| 504 | verifyTmaDescriptorWithMemref(op: *this, descType: getTensorMap().getType()); |
| 505 | if (error.has_value()) |
| 506 | return error.value(); |
| 507 | |
| 508 | if (getTensorMap().getType().getSwizzle() != |
| 509 | TensorMapSwizzleKind::SWIZZLE_128B) { |
| 510 | return emitError() << "supports only " |
| 511 | << stringifyTensorMapSwizzleKind( |
| 512 | TensorMapSwizzleKind::SWIZZLE_128B) |
| 513 | << " is supported for the time being" ; |
| 514 | } |
| 515 | |
| 516 | if (getTensorMap().getType().getInterleave() != |
| 517 | TensorMapInterleaveKind::INTERLEAVE_NONE) { |
| 518 | return emitError() << "supports only " |
| 519 | << stringifyTensorMapInterleaveKind( |
| 520 | TensorMapInterleaveKind::INTERLEAVE_NONE) |
| 521 | << " is supported for the time being" ; |
| 522 | } |
| 523 | |
| 524 | return success(); |
| 525 | } |
| 526 | |
| 527 | //===----------------------------------------------------------------------===// |
| 528 | // WarpgroupMmaOp |
| 529 | //===----------------------------------------------------------------------===// |
| 530 | |
| 531 | LogicalResult isAllowedWGMMADataType(Type typeD, Type typeA, Type typeB) { |
| 532 | // F32 += F16 + F16 |
| 533 | // F16 += F16 + F16 |
| 534 | if (typeA.isF16() && typeB.isF16() && (typeD.isF32() || typeD.isF16())) |
| 535 | return success(); |
| 536 | // F32 += TF32 + TF32 |
| 537 | if (typeA.isTF32() && typeD.isF32() && typeB.isTF32()) |
| 538 | return success(); |
| 539 | // s32 += i8 + i8 |
| 540 | if (typeA.isInteger(width: 8) && typeB.isInteger(width: 8) && typeD.isInteger(width: 32)) |
| 541 | return success(); |
| 542 | // s32 += i1 + i1 |
| 543 | if (typeA.isInteger(width: 1) && typeB.isInteger(width: 1) && typeD.isInteger(width: 32)) |
| 544 | return success(); |
| 545 | // F32 += BF16 + BF16 |
| 546 | // F16 += BF16 + BF16 |
| 547 | if (typeA.isBF16() && typeB.isBF16() && (typeD.isF32() || typeD.isF16())) |
| 548 | return success(); |
| 549 | // F16 += f8 + f8 |
| 550 | // F32 += f8 + f8 |
| 551 | if (isa<Float8E5M2Type, Float8E4M3FNType>(Val: typeA) && |
| 552 | isa<Float8E5M2Type, Float8E4M3FNType>(Val: typeB) && |
| 553 | (typeD.isF32() || typeD.isF16())) |
| 554 | return success(); |
| 555 | |
| 556 | return failure(); |
| 557 | } |
| 558 | |
| 559 | LogicalResult isAllowedSizeM(int sizeM) { |
| 560 | if (sizeM % kWgmmaSizeM) |
| 561 | return failure(); |
| 562 | return success(); |
| 563 | } |
| 564 | |
| 565 | LogicalResult isAllowedSizeN(int sizeN, Type typeA) { |
| 566 | SmallVector<int> allowedN = {8, 16, 24, 32, 40, 48, 56, 64, |
| 567 | 72, 80, 88, 96, 104, 112, 120, 128, |
| 568 | 136, 144, 152, 160, 168, 176, 184, 192, |
| 569 | 200, 208, 216, 224, 232, 240, 248, 256}; |
| 570 | SmallVector<int> allowedNshort = {8, 16, 24, 32, 48, 64, |
| 571 | 80, 96, 112, 128, 144, 160, |
| 572 | 176, 192, 208, 224, 240, 256}; |
| 573 | if (typeA.isBF16() || typeA.isF16() || typeA.isF32() || typeA.isTF32() || |
| 574 | isa<Float8E5M2Type, Float8E4M3FNType>(Val: typeA)) |
| 575 | if (llvm::is_contained(Range&: allowedN, Element: sizeN)) |
| 576 | return success(); |
| 577 | |
| 578 | if (typeA.isInteger(width: 8) || typeA.isInteger(width: 1)) |
| 579 | if (llvm::is_contained(Range&: allowedNshort, Element: sizeN)) |
| 580 | return success(); |
| 581 | return failure(); |
| 582 | } |
| 583 | |
| 584 | LogicalResult WarpgroupMmaOp::verify() { |
| 585 | if (getTransposeA() && !getTransposeB()) |
| 586 | return emitOpError() |
| 587 | << "supports non-transpose A (Row Major) " |
| 588 | "and transpose B (Column Major) for the time being " ; |
| 589 | MemRefType matrixA = getDescriptorA().getType().getTensor(); |
| 590 | MemRefType matrixB = getDescriptorB().getType().getTensor(); |
| 591 | VectorType matrixC = getMatrixC().getType().getFragmented(); |
| 592 | VectorType matrixD = getMatrixD().getType().getFragmented(); |
| 593 | |
| 594 | if (matrixC != matrixD) |
| 595 | return emitOpError() << "type of matrix C and matrix D must be the same" ; |
| 596 | |
| 597 | if (matrixA.getRank() != 2 || matrixB.getRank() != 2 || |
| 598 | matrixC.getRank() != 2 || matrixD.getRank() != 2) { |
| 599 | return emitOpError() |
| 600 | << "has matrices A, B, C and D, they must be 2 dimensional" ; |
| 601 | } |
| 602 | |
| 603 | if (matrixA.getShape()[1] != matrixB.getShape()[0]) |
| 604 | return emitOpError() << "2nd dim matrix-A (" << matrixA.getShape()[1] |
| 605 | << ")!= 1st dim matrix-B (" << matrixB.getShape()[0] |
| 606 | << " )" ; |
| 607 | if (matrixA.getShape()[0] != matrixC.getShape()[0]) |
| 608 | return emitOpError() << "1st dim matrix-A ( " << matrixA.getShape()[0] |
| 609 | << " )!= 1st dim matrix-C ( " << matrixC.getShape()[0] |
| 610 | << " )" ; |
| 611 | if (matrixB.getShape()[1] != matrixC.getShape()[1]) |
| 612 | return emitOpError() << "2nd dim matrix-B ( " << matrixB.getShape()[1] |
| 613 | << " ) != 2nd dim matrix-C ( " << matrixC.getShape()[1] |
| 614 | << " )" ; |
| 615 | |
| 616 | if (failed(Result: isAllowedWGMMADataType(typeD: matrixC.getElementType(), |
| 617 | typeA: matrixA.getElementType(), |
| 618 | typeB: matrixB.getElementType()))) |
| 619 | return emitOpError() << matrixC.getElementType() |
| 620 | << " += " << matrixA.getElementType() << " * " |
| 621 | << matrixB.getElementType() |
| 622 | << ", it is not supported." ; |
| 623 | // Check N |
| 624 | if (failed(Result: isAllowedSizeN(sizeN: matrixB.getDimSize(idx: 1), typeA: matrixA.getElementType()))) { |
| 625 | return emitOpError() << "has input type " << matrixB << " n is set to " |
| 626 | << matrixB.getDimSize(idx: 1) << ", it is not supported" ; |
| 627 | } |
| 628 | |
| 629 | // Currently, f16/bf16 supported |
| 630 | if (!matrixC.getElementType().isF32() && !matrixA.getElementType().isF16() && |
| 631 | !matrixA.getElementType().isBF16()) { |
| 632 | return emitOpError() << "hit a limitation: " << matrixC.getElementType() |
| 633 | << " += " << matrixA.getElementType() << " * " |
| 634 | << matrixB.getElementType() |
| 635 | << ", it is not supported yet" ; |
| 636 | } |
| 637 | |
| 638 | return success(); |
| 639 | } |
| 640 | |
| 641 | LogicalResult WarpgroupMmaStoreOp::verify() { |
| 642 | MemRefType dstMemrefType = getDstMemref().getType(); |
| 643 | VectorType vtype = getMatrixD().getType().getFragmented(); |
| 644 | |
| 645 | // Limitation |
| 646 | if (!vtype.getElementType().isF32()) { |
| 647 | return emitOpError() |
| 648 | << "hit a limitation: only f32 results for the time being" ; |
| 649 | } |
| 650 | if (vtype.getDimSize(idx: 0) != dstMemrefType.getDimSize(idx: 0) || |
| 651 | vtype.getDimSize(idx: 1) != dstMemrefType.getDimSize(idx: 1)) { |
| 652 | return emitOpError() << "results [" << vtype << "][" << vtype.getDimSize(idx: 1) |
| 653 | << "] values. However, destination memref[" |
| 654 | << dstMemrefType.getDimSize(idx: 0) << "][" |
| 655 | << dstMemrefType.getDimSize(idx: 1) |
| 656 | << "] does not have same size as results" ; |
| 657 | } |
| 658 | return success(); |
| 659 | } |
| 660 | |
| 661 | //===----------------------------------------------------------------------===// |
| 662 | // WarpgroupMmaInitAccumulatorOp |
| 663 | //===----------------------------------------------------------------------===// |
| 664 | |
| 665 | LogicalResult WarpgroupMmaInitAccumulatorOp::verify() { |
| 666 | WarpgroupAccumulatorType accType = getMatrixC().getType(); |
| 667 | int64_t sizeM = accType.getFragmented().getDimSize(idx: 0); |
| 668 | int64_t sizeN = accType.getFragmented().getDimSize(idx: 1); |
| 669 | Type elemType = accType.getFragmented().getElementType(); |
| 670 | |
| 671 | if (failed(Result: isAllowedSizeM(sizeM)) || |
| 672 | failed(Result: isAllowedSizeN(sizeN, typeA: elemType))) { |
| 673 | return emitOpError() << "has type " << accType.getFragmented() |
| 674 | << ". It does not fit into warp-group " |
| 675 | "level (wgmma) matrix multiplication instruction " |
| 676 | "(or not supported yet)" ; |
| 677 | } |
| 678 | return success(); |
| 679 | } |
| 680 | |
| 681 | //===----------------------------------------------------------------------===// |
| 682 | // RcpOp |
| 683 | //===----------------------------------------------------------------------===// |
| 684 | |
| 685 | LogicalResult RcpOp::verify() { |
| 686 | bool ftz = getFtz(); |
| 687 | bool approx = getApprox(); |
| 688 | mlir::NVVM::FPRoundingModeAttr rnd = getRoundingAttr(); |
| 689 | // Currently, only `rcp_approx` and `ftz` is supported. |
| 690 | if (!approx || !ftz) { |
| 691 | return emitOpError() |
| 692 | << "has a limitation. non-approx or non-ftz is not supported yet." ; |
| 693 | } |
| 694 | if (rnd.getValue() != mlir::NVVM::FPRoundingMode::NONE) { |
| 695 | return emitOpError() << "has a limitation. " << rnd |
| 696 | << " is not supported yet." ; |
| 697 | } |
| 698 | return success(); |
| 699 | } |
| 700 | |
| 701 | //===----------------------------------------------------------------------===// |
| 702 | // NVGPU_TruncfOp |
| 703 | //===----------------------------------------------------------------------===// |
| 704 | |
| 705 | static LogicalResult verifyConversionShapes(Operation *op, Type inType, |
| 706 | Type outType) { |
| 707 | bool srcIsVector = llvm::isa<VectorType>(Val: inType); |
| 708 | bool dstIsVector = llvm::isa<VectorType>(Val: outType); |
| 709 | if (srcIsVector != dstIsVector) |
| 710 | return op->emitOpError(message: "input and output must both be scalars or both be " |
| 711 | "vectors, got " ) |
| 712 | << inType << " and " << outType; |
| 713 | if (srcIsVector) { |
| 714 | auto srcVector = llvm::cast<VectorType>(Val&: inType); |
| 715 | auto dstVector = llvm::cast<VectorType>(Val&: outType); |
| 716 | if (srcVector.getShape() != dstVector.getShape()) |
| 717 | return op->emitOpError(message: "input and output shapes must match, got " ) |
| 718 | << inType << " and " << outType; |
| 719 | } |
| 720 | return success(); |
| 721 | } |
| 722 | |
| 723 | LogicalResult TruncfOp::verify() { |
| 724 | Type inType = getIn().getType(); |
| 725 | Type outType = getType(); |
| 726 | Type srcType = getElementTypeOrSelf(type: inType); |
| 727 | Type dstType = getElementTypeOrSelf(type: outType); |
| 728 | int srcBitWidth = srcType.getIntOrFloatBitWidth(); |
| 729 | int dstBitWidth = dstType.getIntOrFloatBitWidth(); |
| 730 | auto rnd = getRnd(); |
| 731 | |
| 732 | if (auto result = verifyConversionShapes(op: getOperation(), inType, outType); |
| 733 | failed(Result: result)) |
| 734 | return result; |
| 735 | |
| 736 | if (srcBitWidth <= dstBitWidth) |
| 737 | return emitOpError(message: "result type " ) |
| 738 | << dstType << " must be narrower than operand type " << srcType; |
| 739 | |
| 740 | if (!(srcBitWidth == 64 || srcBitWidth == 32 || srcBitWidth == 16)) |
| 741 | return emitOpError(message: "input type must be 64/32/16 bitwidth, but got " ) |
| 742 | << srcBitWidth; |
| 743 | |
| 744 | if (llvm::isa<Float8E8M0FNUType>(Val: dstType)) { |
| 745 | if (rnd != mlir::NVVM::FPRoundingMode::RZ && |
| 746 | rnd != mlir::NVVM::FPRoundingMode::RP) |
| 747 | return emitOpError(message: "expects RZ or RP rounding mode when result type is " |
| 748 | "e8m0, but got " ) |
| 749 | << getRndAttr(); |
| 750 | } else if (rnd == mlir::NVVM::FPRoundingMode::RS) { |
| 751 | // TODO: Currently, we only support conversions which fit into a single i32 |
| 752 | // register. Support f32->f8/f6/f4 conversions with RS rounding. |
| 753 | if (!(srcBitWidth == 32 && dstBitWidth == 16)) |
| 754 | return emitOpError(message: "RS (stochastic) rounding is only supported for " |
| 755 | "f32->f16/bf16, got " ) |
| 756 | << srcType << " -> " << dstType; |
| 757 | if (!getRandomBits()) |
| 758 | return emitOpError(message: "random_bits operand is required with RS rounding" ); |
| 759 | } else if (srcType.isF64() && dstBitWidth >= 16) { |
| 760 | if (rnd != mlir::NVVM::FPRoundingMode::RN) |
| 761 | return emitOpError(message: "expects RN rounding mode for f64 input, but got " ) |
| 762 | << getRndAttr(); |
| 763 | } else if (srcBitWidth == 32 && dstBitWidth == 16) { |
| 764 | if (rnd != mlir::NVVM::FPRoundingMode::RN && |
| 765 | rnd != mlir::NVVM::FPRoundingMode::RZ) |
| 766 | return emitOpError(message: "expects RN or RZ rounding mode for f32 to f16/bf16, " |
| 767 | "but got " ) |
| 768 | << getRndAttr(); |
| 769 | } else if (rnd != mlir::NVVM::FPRoundingMode::RN) { |
| 770 | return emitOpError(message: "expects RN rounding mode, but got " ) << getRndAttr(); |
| 771 | } |
| 772 | |
| 773 | if (getRandomBits() && rnd != mlir::NVVM::FPRoundingMode::RS) |
| 774 | return emitOpError(message: "random_bits can only be used with RS rounding mode" ); |
| 775 | |
| 776 | return success(); |
| 777 | } |
| 778 | |
| 779 | //===----------------------------------------------------------------------===// |
| 780 | // NVGPU_ExtfOp |
| 781 | //===----------------------------------------------------------------------===// |
| 782 | |
| 783 | LogicalResult ExtfOp::verify() { |
| 784 | Type inType = getIn().getType(); |
| 785 | Type outType = getType(); |
| 786 | Type srcType = getElementTypeOrSelf(type: inType); |
| 787 | Type dstType = getElementTypeOrSelf(type: outType); |
| 788 | int srcBitWidth = srcType.getIntOrFloatBitWidth(); |
| 789 | int dstBitWidth = dstType.getIntOrFloatBitWidth(); |
| 790 | auto rnd = getRnd(); |
| 791 | |
| 792 | if (failed(Result: verifyConversionShapes(op: getOperation(), inType, outType))) |
| 793 | return failure(); |
| 794 | |
| 795 | if (srcBitWidth >= dstBitWidth) |
| 796 | return emitOpError(message: "result type " ) |
| 797 | << dstType << " must be wider than operand type " << srcType; |
| 798 | |
| 799 | if (dstBitWidth != 16 && dstBitWidth != 32 && dstBitWidth != 64) |
| 800 | return emitOpError(message: "result type must be 16, 32, or 64 bitwidth, but got " ) |
| 801 | << dstBitWidth; |
| 802 | |
| 803 | if (llvm::isa<Float8E8M0FNUType>(Val: srcType) && |
| 804 | !llvm::isa<BFloat16Type>(Val: dstType) && !dstType.isF32()) |
| 805 | return emitOpError(message: "expects bf16 or f32 output type when input type is " |
| 806 | "e8m0." ); |
| 807 | |
| 808 | if (rnd != mlir::NVVM::FPRoundingMode::RN) |
| 809 | return emitOpError(message: "expects RN rounding mode, but got " ) << getRndAttr(); |
| 810 | |
| 811 | if (getRelu() && llvm::isa<BFloat16Type>(Val: dstType)) |
| 812 | return emitOpError(message: "relu is not supported for bf16 destination" ); |
| 813 | |
| 814 | return success(); |
| 815 | } |
| 816 | |
| 817 | //===----------------------------------------------------------------------===// |
| 818 | // TableGen'd dialect, type, and op definitions |
| 819 | //===----------------------------------------------------------------------===// |
| 820 | |
| 821 | #define GET_ATTRDEF_CLASSES |
| 822 | #include "mlir/Dialect/NVGPU/IR/NVGPUAttrDefs.cpp.inc" |
| 823 | |
| 824 | #include "mlir/Dialect/NVGPU/IR/NVGPUEnums.cpp.inc" |
| 825 | |
| 826 | #define GET_OP_CLASSES |
| 827 | #include "mlir/Dialect/NVGPU/IR/NVGPUOps.cpp.inc" |
| 828 | |
| 829 | #define GET_TYPEDEF_CLASSES |
| 830 | #include "mlir/Dialect/NVGPU/IR/NVGPUTypeDefs.cpp.inc" |
| 831 | |