Add reverse AD for llvm.intr.memcpy #2841
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -17,7 +17,9 @@ | |
| #include "Interfaces/AutoDiffTypeInterface.h" | ||
| #include "Interfaces/GradientUtils.h" | ||
| #include "Interfaces/GradientUtilsReverse.h" | ||
| #include "mlir/Dialect/Arith/IR/Arith.h" | ||
| #include "mlir/Dialect/LLVMIR/LLVMDialect.h" | ||
| #include "mlir/Dialect/SCF/IR/SCF.h" | ||
| #include "mlir/IR/DialectRegistry.h" | ||
| #include "mlir/Support/LogicalResult.h" | ||
|
|
||
|
|
@@ -297,6 +299,109 @@ struct InsertValueOpInterfaceReverse | |
| MGradientUtilsReverse *gutils) const {} | ||
| }; | ||
|
|
||
| struct MemcpyOpInterfaceReverse | ||
| : public ReverseAutoDiffOpInterface::ExternalModel<MemcpyOpInterfaceReverse, | ||
| LLVM::MemcpyOp> { | ||
|
|
||
| static Type inferElemType(LLVM::MemcpyOp cp) { | ||
| if (auto t = cp->getAttrOfType<TypeAttr>("enzyme.elem_type")) | ||
| return t.getValue(); | ||
| auto walk = [](Value p) -> Type { | ||
| for (Operation *user : p.getUsers()) { | ||
| if (auto ld = dyn_cast<LLVM::LoadOp>(user)) | ||
| if (isa<AutoDiffTypeInterface>(ld.getType())) | ||
| return ld.getType(); | ||
| if (auto st = dyn_cast<LLVM::StoreOp>(user)) | ||
| if (isa<AutoDiffTypeInterface>(st.getValue().getType())) | ||
| return st.getValue().getType(); | ||
| } | ||
| return nullptr; | ||
| }; | ||
| if (Type t = walk(cp.getDst())) | ||
| return t; | ||
| if (Type t = walk(cp.getSrc())) | ||
| return t; | ||
| return Float64Type::get(cp.getContext()); | ||
| } | ||
|
|
||
| SmallVector<Value> cacheValues(Operation *op, | ||
| MGradientUtilsReverse *gutils) const { | ||
| auto cp = cast<LLVM::MemcpyOp>(op); | ||
| if (gutils->isConstantValue(cp.getDst())) | ||
| return {}; | ||
| bool srcActive = !gutils->isConstantValue(cp.getSrc()); | ||
| OpBuilder cb(gutils->getNewFromOriginal(op)); | ||
| SmallVector<Value> caches; | ||
| caches.push_back( | ||
| gutils->initAndPushCache(gutils->invertPointerM(cp.getDst(), cb), cb)); | ||
| caches.push_back(gutils->initAndPushCache( | ||
| srcActive ? gutils->invertPointerM(cp.getSrc(), cb) | ||
| : gutils->getNewFromOriginal(cp.getSrc()), | ||
| cb)); | ||
| caches.push_back( | ||
| gutils->initAndPushCache(gutils->getNewFromOriginal(cp.getLen()), cb)); | ||
| return caches; | ||
| } | ||
|
|
||
| void createShadowValues(Operation *op, OpBuilder &builder, | ||
| MGradientUtilsReverse *gutils) const {} | ||
|
|
||
| LogicalResult createReverseModeAdjoint(Operation *op, OpBuilder &builder, | ||
| MGradientUtilsReverse *gutils, | ||
| SmallVector<Value> caches) const { | ||
| auto cp = cast<LLVM::MemcpyOp>(op); | ||
| if (gutils->isConstantValue(cp.getDst())) | ||
| return success(); | ||
| bool srcActive = !gutils->isConstantValue(cp.getSrc()); | ||
|
|
||
| Value dDst = gutils->popCache(caches[0], builder); | ||
| Value dSrc = gutils->popCache(caches[1], builder); | ||
| Value len = gutils->popCache(caches[2], builder); | ||
|
|
||
| Type elemTy = inferElemType(cp); | ||
| auto adt = dyn_cast<AutoDiffTypeInterface>(elemTy); | ||
| if (!adt || !elemTy.isIntOrFloat()) | ||
| return op->emitError() | ||
| << "memcpy reverse: unsupported element type " << elemTy | ||
| << " (annotate enzyme.elem_type or lower to scalar stores)"; | ||
|
|
||
| Location loc = op->getLoc(); | ||
| unsigned bytes = (elemTy.getIntOrFloatBitWidth() + 7) / 8; | ||
|
|
||
| // n_elements = len / sizeof(elemTy) | ||
| Value byteSz = | ||
| LLVM::ConstantOp::create(builder, loc, len.getType(), | ||
| builder.getIntegerAttr(len.getType(), bytes)); | ||
| Value nInt = LLVM::SDivOp::create(builder, loc, len, byteSz); | ||
|
There was a problem hiding this comment.
Useful? React with 👍 / 👎. |
||
| Value n = | ||
| arith::IndexCastOp::create(builder, loc, builder.getIndexType(), nInt); | ||
|
|
||
| Value c0 = arith::ConstantIndexOp::create(builder, loc, 0); | ||
| Value c1 = arith::ConstantIndexOp::create(builder, loc, 1); | ||
| Value zeroElem = adt.createNullValue(builder, loc); | ||
| Type ptrTy = cp.getDst().getType(); | ||
|
|
||
| auto forOp = scf::ForOp::create(builder, loc, c0, n, c1); | ||
| OpBuilder body(forOp.getBody()->getTerminator()); | ||
|
Collaborator
There was a problem hiding this comment. better to use the provided
Collaborator
Author
There was a problem hiding this comment. @Pangoraw thanks, I have modified
we have covert-polygeist-to-llvm in the downstream, I guess it may could work for it
Collaborator
There was a problem hiding this comment. ah yes perfect in this case |
||
| Value ivIdx = forOp.getInductionVar(); | ||
| Value iv = arith::IndexCastOp::create(body, loc, len.getType(), ivIdx); | ||
|
|
||
| Value gDst = LLVM::GEPOp::create(body, loc, ptrTy, elemTy, dDst, | ||
| ArrayRef<LLVM::GEPArg>{iv}); | ||
| Value vDst = LLVM::LoadOp::create(body, loc, elemTy, gDst); | ||
| if (srcActive) { | ||
| Value gSrc = LLVM::GEPOp::create(body, loc, ptrTy, elemTy, dSrc, | ||
| ArrayRef<LLVM::GEPArg>{iv}); | ||
|
Comment on lines
+413
to
+414
There was a problem hiding this comment.
When a memcpy copies between different LLVM pointer address spaces, Useful? React with 👍 / 👎. |
||
| Value vSrc = LLVM::LoadOp::create(body, loc, elemTy, gSrc); | ||
| Value sum = adt.createAddOp(body, loc, vSrc, vDst); | ||
| LLVM::StoreOp::create(body, loc, sum, gSrc); | ||
| } | ||
| LLVM::StoreOp::create(body, loc, zeroElem, gDst); | ||
|
|
||
| return success(); | ||
| } | ||
| }; | ||
|
|
||
| std::optional<Value> findPtrSize(Value ptr) { | ||
| if (auto allocOp = ptr.getDefiningOp<llvm_ext::AllocOp>()) | ||
| return allocOp.getSize(); | ||
|
|
@@ -467,6 +572,7 @@ void mlir::enzyme::registerLLVMDialectAutoDiffInterface( | |
| *context); | ||
| LLVM::InsertValueOp::attachInterface<InsertValueOpInterfaceReverse>( | ||
| *context); | ||
| LLVM::MemcpyOp::attachInterface<MemcpyOpInterfaceReverse>(*context); | ||
| LLVM::UnreachableOp::template attachInterface< | ||
| detail::NoopRevAutoDiffInterface<LLVM::UnreachableOp>>(*context); | ||
| LLVM::LLVMFuncOp::attachInterface<AutoDiffLLVMFuncOpFunctionInterface>( | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,64 @@ | ||
| // RUN: %eopt --enzyme %s | FileCheck %s | ||
|
|
||
| func.func @copy1(%dst: !llvm.ptr, %src: !llvm.ptr, %n: i64) { | ||
| "llvm.intr.memcpy"(%dst, %src, %n) | ||
| <{arg_attrs = [{llvm.align = 8 : i64}], isVolatile = false}> | ||
| : (!llvm.ptr, !llvm.ptr, i64) -> () | ||
| return | ||
| } | ||
|
|
||
| func.func @dcopy1(%dst: !llvm.ptr, %ddst: !llvm.ptr, | ||
| %src: !llvm.ptr, %dsrc: !llvm.ptr, %n: i64) { | ||
| enzyme.autodiff @copy1(%dst, %ddst, %src, %dsrc, %n) { | ||
| activity = [#enzyme<activity enzyme_dup>, | ||
| #enzyme<activity enzyme_dup>, | ||
| #enzyme<activity enzyme_const>], | ||
| ret_activity = [] | ||
| } : (!llvm.ptr, !llvm.ptr, !llvm.ptr, !llvm.ptr, i64) -> () | ||
| return | ||
| } | ||
|
|
||
| func.func @copy2(%dst: !llvm.ptr, %src: !llvm.ptr, %n: i64) { | ||
| "llvm.intr.memcpy"(%dst, %src, %n) | ||
| <{arg_attrs = [{llvm.align = 8 : i64}, {llvm.align = 8 : i64}, {}], | ||
| isVolatile = false}> | ||
| : (!llvm.ptr, !llvm.ptr, i64) -> () | ||
| return | ||
| } | ||
|
|
||
| func.func @dcopy2(%dst: !llvm.ptr, %ddst: !llvm.ptr, | ||
| %src: !llvm.ptr, %dsrc: !llvm.ptr, %n: i64) { | ||
| enzyme.autodiff @copy2(%dst, %ddst, %src, %dsrc, %n) { | ||
| activity = [#enzyme<activity enzyme_dup>, | ||
| #enzyme<activity enzyme_dup>, | ||
| #enzyme<activity enzyme_const>], | ||
| ret_activity = [] | ||
| } : (!llvm.ptr, !llvm.ptr, !llvm.ptr, !llvm.ptr, i64) -> () | ||
| return | ||
| } | ||
|
|
||
| // CHECK-LABEL: func.func private @diffecopy1( | ||
| // Forward: the primal memcpy is preserved. | ||
| // CHECK: "llvm.intr.memcpy" | ||
| // Reverse: n / sizeof(f64) element-wise loop, d_src[i] += d_dst[i]; d_dst[i]=0. | ||
| // CHECK: %[[BYTES:.+]] = llvm.mlir.constant(8 : i64) : i64 | ||
| // CHECK: llvm.sdiv %{{.+}}, %[[BYTES]] : i64 | ||
| // CHECK: arith.index_cast | ||
| // CHECK: %[[ZERO:.+]] = arith.constant 0.000000e+00 : f64 | ||
| // CHECK: scf.for | ||
| // CHECK: llvm.getelementptr %{{.+}}[%{{.+}}] : (!llvm.ptr, i64) -> !llvm.ptr, f64 | ||
| // CHECK: llvm.load %{{.+}} : !llvm.ptr -> f64 | ||
| // CHECK: llvm.getelementptr %{{.+}}[%{{.+}}] : (!llvm.ptr, i64) -> !llvm.ptr, f64 | ||
| // CHECK: llvm.load %{{.+}} : !llvm.ptr -> f64 | ||
| // CHECK: %[[SUM:.+]] = arith.addf | ||
| // CHECK: llvm.store %[[SUM]], %{{.+}} : f64, !llvm.ptr | ||
| // CHECK: llvm.store %[[ZERO]], %{{.+}} : f64, !llvm.ptr | ||
|
|
||
| // CHECK-LABEL: func.func private @diffecopy2( | ||
| // CHECK: "llvm.intr.memcpy" | ||
| // CHECK: llvm.mlir.constant(8 : i64) : i64 | ||
| // CHECK: %[[ZERO2:.+]] = arith.constant 0.000000e+00 : f64 | ||
| // CHECK: scf.for | ||
| // CHECK: %[[SUM2:.+]] = arith.addf | ||
| // CHECK: llvm.store %[[SUM2]], %{{.+}} : f64, !llvm.ptr | ||
| // CHECK: llvm.store %[[ZERO2]], %{{.+}} : f64, !llvm.ptr |
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this is wrong, we should throw an error if things can't be deduced. This is also somewhat hacky and we should use type analysis
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ok, I will fix it
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@wsmoses I have fixed it, but only for the case where the inferred element type is a scalar int or float and using upstream MLIR for inferring types. The memcpy ops in RSBench that copy arrays of
struct<(f64, f64)>andarray<10 x f64>need to handle it recursively, I plan to open a new PR for it.