2024-07-11 02:04:17 +08:00
|
|
|
#include "torch-mlir/Conversion/TorchOnnxToTorch/Patterns.h"
|
|
|
|
#include "torch-mlir/Conversion/TorchOnnxToTorch/Utils.h"
|
|
|
|
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
|
|
|
|
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
|
|
|
|
|
|
|
|
using namespace mlir;
|
|
|
|
using namespace mlir::torch::Torch;
|
|
|
|
|
|
|
|
namespace mlir::torch::onnx_c {
|
|
|
|
|
|
|
|
/**
|
|
|
|
* @brief Splits the input tensor based on the provided direction.
|
|
|
|
*
|
|
|
|
* This function is used to split the LSTM parameters (W, R, B) into forward
|
|
|
|
* and backward directions. The input tensor is expected to have the forward
|
|
|
|
* and backward parameters concatenated along the 0th dimension. The function
|
|
|
|
* returns a tensor that contains the parameters for the specified direction.
|
|
|
|
*
|
|
|
|
* @param direction The direction to split out. 0 for forward, 1 for backward.
|
|
|
|
* @param input The input tensor to split.
|
|
|
|
* @return The split tensor for the specified direction.
|
|
|
|
*/
|
|
|
|
Value getDirection(ImplicitLocOpBuilder b, int64_t direction, Value input) {
|
|
|
|
auto inputType = cast<ValueTensorType>(input.getType());
|
|
|
|
auto outputType = cast<ValueTensorType>(inputType.getWithSizesAndDtype(
|
|
|
|
llvm::SmallVector<int64_t>{inputType.getSizes().drop_front()},
|
|
|
|
inputType.getDtype()));
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value selectDim = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstDirection =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(direction));
|
|
|
|
return b.create<AtenSelectIntOp>(outputType, input, selectDim, cstDirection);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct RnnWeights {
|
|
|
|
Value Wi;
|
|
|
|
Value Ri;
|
|
|
|
Value Wbi;
|
|
|
|
Value Rbi;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct RnnActivations {
|
|
|
|
std::string f;
|
|
|
|
};
|
|
|
|
|
|
|
|
Value rnn_cell(ImplicitLocOpBuilder &b, Value Xt, Value H_prev,
|
|
|
|
RnnWeights weights, RnnActivations activations) {
|
|
|
|
auto hTy = cast<ValueTensorType>(H_prev.getType());
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
|
|
|
|
Value i_x = b.create<AtenLinearOp>(hTy, Xt, weights.Wi, weights.Wbi);
|
|
|
|
Value i_h = b.create<AtenLinearOp>(hTy, H_prev, weights.Ri, weights.Rbi);
|
|
|
|
Value i = b.create<AtenAddTensorOp>(hTy, i_x, i_h, cstOne);
|
|
|
|
|
|
|
|
Value H_new = createActivationByName(b, activations.f, i);
|
|
|
|
return H_new;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct RnnLayerOutput {
|
|
|
|
Value Y;
|
|
|
|
Value Y_h;
|
|
|
|
};
|
|
|
|
|
|
|
|
RnnLayerOutput rnn_layer(ImplicitLocOpBuilder &b, Value X, Value initial_h,
|
|
|
|
RnnWeights weights, RnnActivations activations) {
|
|
|
|
Location loc = b.getLoc();
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto hTy = cast<ValueTensorType>(initial_h.getType());
|
|
|
|
int64_t seq_len = xTy.getSizes()[0];
|
|
|
|
int64_t batch_size = xTy.getSizes()[1];
|
|
|
|
int64_t input_size = xTy.getSizes()[2];
|
|
|
|
int64_t hidden_size = hTy.getSizes()[1];
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
Value cstSeqLen =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(seq_len));
|
|
|
|
Value cstBatchSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size));
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
|
|
|
|
auto yTy = b.getType<ValueTensorType>(
|
|
|
|
SmallVector<int64_t>{seq_len, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
|
|
|
|
auto YShapeList = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType),
|
|
|
|
ValueRange({cstSeqLen, cstBatchSize, cstHiddenSize}));
|
|
|
|
|
|
|
|
int64_t hDtypeInt =
|
|
|
|
static_cast<int64_t>(getScalarTypeForType(hTy.getDtype()));
|
|
|
|
Value hDtypeIntVal =
|
|
|
|
b.create<ConstantIntOp>(loc, b.getI64IntegerAttr(hDtypeInt));
|
|
|
|
|
|
|
|
Value Y_initial = b.create<AtenZerosOp>(yTy, YShapeList, hDtypeIntVal,
|
|
|
|
cstNone, cstNone, cstNone);
|
|
|
|
|
|
|
|
Value maxTripCount =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(seq_len));
|
|
|
|
Value loopConditionTrue = b.create<ConstantBoolOp>(true);
|
|
|
|
|
|
|
|
Type loopIndexType = intType;
|
|
|
|
auto loop = b.create<PrimLoopOp>(TypeRange({yTy, hTy}), maxTripCount,
|
|
|
|
loopConditionTrue,
|
|
|
|
ValueRange({Y_initial, initial_h}));
|
|
|
|
{
|
|
|
|
OpBuilder::InsertionGuard guard(b);
|
|
|
|
Block *loopBody =
|
|
|
|
b.createBlock(&loop.getRegion(), loop.getRegion().begin(),
|
|
|
|
TypeRange({
|
|
|
|
loopIndexType,
|
|
|
|
yTy,
|
|
|
|
hTy,
|
|
|
|
}),
|
|
|
|
{loc, loc, loc} // locs for the loop body arguments
|
|
|
|
);
|
|
|
|
|
|
|
|
Value loopIndex = loopBody->getArgument(0);
|
|
|
|
Value Y_prev = loopBody->getArgument(1);
|
|
|
|
Value H_prev = loopBody->getArgument(2);
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto XtType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{batch_size, input_size}, xTy.getDtype());
|
|
|
|
|
|
|
|
Value Xt = b.create<AtenSelectIntOp>(XtType, X, cstZero, loopIndex);
|
|
|
|
|
|
|
|
Value H_new = rnn_cell(b, Xt, H_prev, weights, activations);
|
|
|
|
|
|
|
|
Type hTyUnsqueezed = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{1, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
Value H_new_unsqueezed =
|
|
|
|
b.create<AtenUnsqueezeOp>(hTyUnsqueezed, H_new, cstZero);
|
|
|
|
|
|
|
|
auto loopIndexPlusOne = b.create<AtenAddIntOp>(intType, loopIndex, cstOne);
|
|
|
|
Value Y_new =
|
|
|
|
b.create<AtenSliceScatterOp>(yTy, Y_prev, H_new_unsqueezed, cstZero,
|
|
|
|
loopIndex, loopIndexPlusOne, cstOne);
|
|
|
|
|
|
|
|
b.create<PrimLoopConditionOp>(loopConditionTrue,
|
|
|
|
ValueRange({Y_new, H_new}));
|
|
|
|
}
|
|
|
|
RnnLayerOutput output;
|
|
|
|
output.Y = loop.getResult(0);
|
|
|
|
output.Y_h = loop.getResult(1);
|
|
|
|
return output;
|
|
|
|
}
|
2024-08-14 05:34:25 +08:00
|
|
|
|
|
|
|
static Value StaticTranspose(ImplicitLocOpBuilder b, Value value, int64_t dim0,
|
|
|
|
int64_t dim1) {
|
|
|
|
auto valueTy = cast<ValueTensorType>(value.getType());
|
|
|
|
|
|
|
|
SmallVector<int64_t> valueShape(valueTy.getSizes());
|
|
|
|
std::swap(valueShape[dim0], valueShape[dim1]);
|
|
|
|
valueTy = b.getType<ValueTensorType>(valueShape, valueTy.getDtype());
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value dim0v = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(dim0));
|
|
|
|
Value dim1v = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(dim1));
|
|
|
|
|
|
|
|
return b.create<AtenTransposeIntOp>(valueTy, value, dim0v, dim1v);
|
|
|
|
}
|
|
|
|
|
2024-07-11 02:04:17 +08:00
|
|
|
LogicalResult OnnxRnnExpander(OpBinder binder,
|
|
|
|
ConversionPatternRewriter &rewriter) {
|
|
|
|
Location loc = binder.getLoc();
|
|
|
|
mlir::ImplicitLocOpBuilder b(loc, rewriter);
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
|
|
|
|
int64_t num_directions = Torch::kUnknownSize;
|
|
|
|
int64_t hidden_size = Torch::kUnknownSize;
|
|
|
|
|
|
|
|
// Attributes
|
|
|
|
llvm::SmallVector<std::string> activationsList;
|
|
|
|
RnnActivations activations;
|
|
|
|
activations.f = "Tanh";
|
|
|
|
if (!binder.stringArrayAttr(activationsList, "activations") &&
|
|
|
|
activationsList.size() > 0) {
|
|
|
|
if (activationsList.size() == 1) {
|
|
|
|
activations.f = activationsList[0];
|
|
|
|
} else if (activationsList.size() == 2) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Bi-directional RNN is not yet supported, yet two "
|
|
|
|
"activation function names are provided");
|
|
|
|
} else {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Unsupported number of activation functions: " +
|
|
|
|
std::to_string(activationsList.size()) +
|
|
|
|
" are provided.");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string direction;
|
|
|
|
if (!binder.customOpNameStringAttr(direction, "direction", "forward") &&
|
|
|
|
direction != "forward")
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported direction attribute value. "
|
|
|
|
"Only 'forward' is supported but '" +
|
|
|
|
direction + "' is provided.");
|
|
|
|
num_directions = (direction == "bidirectional") ? 2 : 1;
|
|
|
|
|
|
|
|
// hidden_size is required according to the docs,
|
|
|
|
// but if we encounter a model that doesn't have it
|
|
|
|
// that we really want to just push through, consider
|
|
|
|
// deleting this check and making it infer the hidden size
|
|
|
|
if (binder.s64IntegerAttr(hidden_size, "hidden_size"))
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Missing required attribute hidden_size");
|
|
|
|
|
2024-08-14 05:34:25 +08:00
|
|
|
// Other attributes
|
|
|
|
int64_t layout;
|
|
|
|
if (binder.s64IntegerAttr(layout, "layout", 0))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported layout attribute type.");
|
|
|
|
|
|
|
|
if (layout < 0 || layout > 1)
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported layout attribute value.");
|
|
|
|
|
2024-07-11 02:04:17 +08:00
|
|
|
// Result types
|
|
|
|
ValueTensorType yTy, Y_hType;
|
2024-08-17 00:23:47 +08:00
|
|
|
auto hasResult0 = binder.tensorResultTypeAtIndex(yTy, 0);
|
|
|
|
auto hasResult1 = binder.tensorResultTypeAtIndex(Y_hType, 1);
|
|
|
|
|
|
|
|
if (hasResult0 && hasResult1) {
|
2024-07-11 02:04:17 +08:00
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"At least one output must be present");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Inputs
|
|
|
|
Value X, W, R, B, initial_h;
|
|
|
|
if (binder.tensorOperandAtIndex(X, 0))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor X");
|
|
|
|
if (binder.tensorOperandAtIndex(W, 1))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor W");
|
|
|
|
if (binder.tensorOperandAtIndex(R, 2))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor R");
|
|
|
|
if (binder.tensorOperandAtIndex(B, 3)) {
|
|
|
|
// if no b found, set to null and create one later
|
|
|
|
B = nullptr;
|
|
|
|
}
|
|
|
|
if (binder.tensorOperandAtIndex(initial_h, 5)) {
|
|
|
|
// if no initial_h found, set to null and create one later
|
|
|
|
initial_h = nullptr;
|
|
|
|
}
|
|
|
|
|
2024-08-14 05:34:25 +08:00
|
|
|
if (layout == 1) {
|
|
|
|
X = StaticTranspose(b, X, 0, 1);
|
|
|
|
if (initial_h)
|
|
|
|
initial_h = StaticTranspose(b, initial_h, 0, 1);
|
|
|
|
}
|
|
|
|
|
2024-07-11 02:04:17 +08:00
|
|
|
// validation
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto wTy = cast<ValueTensorType>(W.getType());
|
|
|
|
auto rTy = cast<ValueTensorType>(R.getType());
|
|
|
|
auto wShape = wTy.getSizes();
|
|
|
|
auto xShape = xTy.getSizes();
|
|
|
|
auto rShape = rTy.getSizes();
|
|
|
|
assert(wShape.size() == 3);
|
|
|
|
|
2024-08-14 05:34:25 +08:00
|
|
|
int64_t seq_len = xShape[0];
|
2024-07-11 02:04:17 +08:00
|
|
|
int64_t batch_size = xShape[1];
|
|
|
|
int64_t x_input_size = xShape[2];
|
|
|
|
|
|
|
|
int64_t w_num_directions = wShape[0];
|
|
|
|
int64_t w_hidden_size = wShape[1];
|
|
|
|
int64_t w_input_size = wShape[2];
|
|
|
|
|
|
|
|
int64_t r_num_directions = rShape[0];
|
|
|
|
if (rShape[1] != rShape[2])
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op,
|
|
|
|
"R tensor must be square, but got shape: " + std::to_string(rShape[1]) +
|
|
|
|
"x" + std::to_string(rShape[2]));
|
|
|
|
int64_t r_hidden_size = rShape[1];
|
|
|
|
|
|
|
|
// validate input size
|
|
|
|
if (x_input_size != w_input_size) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "input_size inferred from shape of X (" +
|
|
|
|
std::to_string(x_input_size) +
|
|
|
|
") does not match the input_size attribute value (" +
|
|
|
|
std::to_string(w_input_size) + ")");
|
|
|
|
}
|
|
|
|
|
|
|
|
// validate hidden size
|
|
|
|
if (w_hidden_size != Torch::kUnknownSize && hidden_size != w_hidden_size) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "hidden_size inferred from shape of W (" +
|
|
|
|
std::to_string(w_hidden_size) +
|
|
|
|
") does not match the hidden_size attribute value (" +
|
|
|
|
std::to_string(hidden_size) + ")");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (r_hidden_size != Torch::kUnknownSize && hidden_size != r_hidden_size) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "hidden_size inferred from shape of R (" +
|
|
|
|
std::to_string(r_hidden_size) +
|
|
|
|
") does not match the hidden_size attribute value (" +
|
|
|
|
std::to_string(hidden_size) + ")");
|
|
|
|
}
|
|
|
|
|
|
|
|
// validate num directions
|
|
|
|
if (w_num_directions != Torch::kUnknownSize &&
|
|
|
|
w_num_directions != num_directions) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "num_directions from shape of W (" +
|
|
|
|
std::to_string(w_num_directions) +
|
|
|
|
") does not match the direction attribute value (" +
|
|
|
|
direction + ")");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (r_num_directions != Torch::kUnknownSize &&
|
|
|
|
r_num_directions != num_directions) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "num_directions from shape of R (" +
|
|
|
|
std::to_string(r_num_directions) +
|
|
|
|
") does not match the direction attribute value (" +
|
|
|
|
direction + ")");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (num_directions != 1) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op,
|
|
|
|
"Unsupported num_directions. Only 1 is currently supported but " +
|
|
|
|
std::to_string(num_directions) + " is provided.");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Create B and initial_h if not provided,
|
|
|
|
// using same dtype as X
|
|
|
|
Value cstXDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
|
|
|
|
if (B == nullptr) {
|
|
|
|
SmallVector<int64_t> BShape = {num_directions, 2 * hidden_size};
|
|
|
|
SmallVector<Value> BShapeListContents = {
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions)),
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2 * hidden_size))};
|
|
|
|
Value BShapeList = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType), BShapeListContents);
|
|
|
|
auto BType = b.getType<ValueTensorType>(BShape, wTy.getDtype());
|
|
|
|
B = b.create<Torch::AtenZerosOp>(BType, BShapeList, cstXDtype, cstNone,
|
|
|
|
cstNone, cstNone);
|
|
|
|
}
|
|
|
|
if (initial_h == nullptr) {
|
|
|
|
SmallVector<int64_t> initial_h_shape = {num_directions, batch_size,
|
|
|
|
hidden_size};
|
|
|
|
SmallVector<Value> initial_h_shape_list_contents = {
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions)),
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size)),
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size))};
|
|
|
|
Value initial_h_shape_list = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType), initial_h_shape_list_contents);
|
|
|
|
auto initial_h_type =
|
|
|
|
b.getType<ValueTensorType>(initial_h_shape, wTy.getDtype());
|
|
|
|
initial_h =
|
|
|
|
b.create<Torch::AtenZerosOp>(initial_h_type, initial_h_shape_list,
|
|
|
|
cstXDtype, cstNone, cstNone, cstNone);
|
|
|
|
}
|
|
|
|
|
|
|
|
Value W_forward = getDirection(b, 0, W);
|
|
|
|
Value R_forward = getDirection(b, 0, R);
|
|
|
|
Value B_forward = getDirection(b, 0, B);
|
|
|
|
Value initial_h_forward = getDirection(b, 0, initial_h);
|
|
|
|
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
|
|
|
|
RnnWeights weights;
|
|
|
|
weights.Wi = W_forward;
|
|
|
|
weights.Ri = R_forward;
|
|
|
|
weights.Wbi = b.create<AtenSliceTensorOp>(
|
|
|
|
b.getType<ValueTensorType>(llvm::SmallVector<int64_t>{hidden_size},
|
|
|
|
wTy.getDtype()),
|
|
|
|
B_forward, cstZero, cstZero, cstHiddenSize, cstOne);
|
|
|
|
weights.Rbi = b.create<AtenSliceTensorOp>(
|
|
|
|
b.getType<ValueTensorType>(llvm::SmallVector<int64_t>{hidden_size},
|
|
|
|
wTy.getDtype()),
|
|
|
|
B_forward, cstZero, cstHiddenSize,
|
|
|
|
b.create<AtenMulIntOp>(
|
|
|
|
cstHiddenSize,
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2))),
|
|
|
|
cstOne);
|
|
|
|
|
|
|
|
RnnLayerOutput rnnLayerOutput =
|
|
|
|
rnn_layer(b, X, initial_h_forward, weights, activations);
|
|
|
|
|
|
|
|
auto Y_h_unsqueezed_type = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{num_directions, batch_size, hidden_size},
|
|
|
|
cast<ValueTensorType>(rnnLayerOutput.Y_h.getType()).getDtype());
|
|
|
|
Value Y_h_unsqueezed = b.create<AtenUnsqueezeOp>(Y_h_unsqueezed_type,
|
|
|
|
rnnLayerOutput.Y_h, cstZero);
|
|
|
|
|
2024-08-14 05:34:25 +08:00
|
|
|
auto Y_unsqueezed_type = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{seq_len, num_directions, batch_size,
|
|
|
|
hidden_size},
|
|
|
|
cast<ValueTensorType>(rnnLayerOutput.Y_h.getType()).getDtype());
|
|
|
|
Value Y_unsqueezed =
|
|
|
|
b.create<AtenUnsqueezeOp>(Y_unsqueezed_type, rnnLayerOutput.Y, cstOne);
|
|
|
|
|
|
|
|
if (layout == 1) {
|
|
|
|
Y_h_unsqueezed = StaticTranspose(b, Y_h_unsqueezed, 0, 1);
|
|
|
|
Y_unsqueezed = StaticTranspose(b, Y_unsqueezed, 1, 2);
|
|
|
|
Y_unsqueezed = StaticTranspose(b, Y_unsqueezed, 0, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!yTy)
|
|
|
|
Y_unsqueezed = cstNone;
|
|
|
|
if (!Y_hType)
|
|
|
|
Y_h_unsqueezed = cstNone;
|
|
|
|
|
2024-07-11 02:04:17 +08:00
|
|
|
rewriter.replaceOp(binder.op, {Y_unsqueezed, Y_h_unsqueezed});
|
|
|
|
return success();
|
|
|
|
}
|
|
|
|
|
|
|
|
// @struct LstmWeights
|
|
|
|
// @brief A structure to hold LSTM weights.
|
|
|
|
//
|
|
|
|
// Each W_ weight matrix should have shape [hidden_size, input_size].
|
|
|
|
// Each R_ weight matrix should have shape [hidden_size, hidden_size].
|
|
|
|
// Each bias vector should have shape [4 * hidden_size].
|
|
|
|
struct LstmWeights {
|
|
|
|
Value W_i, W_o, W_f, W_c;
|
|
|
|
Value R_i, R_o, R_f, R_c;
|
|
|
|
Value Wb_i, Wb_o, Wb_f, Wb_c;
|
|
|
|
Value Rb_i, Rb_o, Rb_f, Rb_c;
|
|
|
|
};
|
|
|
|
struct LstmActivations {
|
|
|
|
std::string f;
|
|
|
|
std::string g;
|
|
|
|
std::string h;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct LstmCellState {
|
|
|
|
Value H;
|
|
|
|
Value C;
|
|
|
|
};
|
|
|
|
// This function represents a Long Short-Term Memory (LSTM) cell operation.
|
|
|
|
//
|
|
|
|
// @param b A builder for constructing operations.
|
|
|
|
// @param Xt The input sequence. It has a shape of [batch_size, input_size].
|
|
|
|
// @param H_prev The previous hidden state. It has a shape of [batch_size,
|
|
|
|
// hidden_size].
|
|
|
|
// @param C_prev The previous cell state. It has a shape of [batch_size,
|
|
|
|
// hidden_size].
|
|
|
|
// @param weights The weights for the LSTM cell. See @ref LstmWeights for shapes
|
|
|
|
// @param activations The activation functions for the LSTM cell. Members f,g,h
|
|
|
|
// correspond to f,g,h in https://onnx.ai/onnx/operators/onnx__LSTM.html
|
|
|
|
// @return The state of the LSTM cell after the operation.
|
|
|
|
LstmCellState lstm_cell(ImplicitLocOpBuilder &b, Value Xt, Value H_prev,
|
|
|
|
Value C_prev, LstmWeights weights,
|
|
|
|
LstmActivations activations) {
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
auto hTy = cast<ValueTensorType>(H_prev.getType());
|
|
|
|
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
|
|
|
|
// Apply linear/matmul for each gate separately
|
|
|
|
// names are consistent with ONNX LSTM documentation
|
|
|
|
Value i_x = b.create<AtenLinearOp>(hTy, Xt, weights.W_i, weights.Wb_i);
|
|
|
|
Value i_h = b.create<AtenLinearOp>(hTy, H_prev, weights.R_i, weights.Rb_i);
|
|
|
|
Value i = b.create<AtenAddTensorOp>(hTy, i_x, i_h, cstOne);
|
|
|
|
Value i_act = createActivationByName(b, activations.f, i);
|
|
|
|
|
|
|
|
Value o_x = b.create<AtenLinearOp>(hTy, Xt, weights.W_o, weights.Wb_o);
|
|
|
|
Value o_h = b.create<AtenLinearOp>(hTy, H_prev, weights.R_o, weights.Rb_o);
|
|
|
|
Value o = b.create<AtenAddTensorOp>(hTy, o_x, o_h, cstOne);
|
|
|
|
Value o_act = createActivationByName(b, activations.f, o);
|
|
|
|
|
|
|
|
Value f_x = b.create<AtenLinearOp>(hTy, Xt, weights.W_f, weights.Wb_f);
|
|
|
|
Value f_h = b.create<AtenLinearOp>(hTy, H_prev, weights.R_f, weights.Rb_f);
|
|
|
|
Value f = b.create<AtenAddTensorOp>(hTy, f_x, f_h, cstOne);
|
|
|
|
Value f_act = createActivationByName(b, activations.f, f);
|
|
|
|
|
|
|
|
Value ct_x = b.create<AtenLinearOp>(hTy, Xt, weights.W_c, weights.Wb_c);
|
|
|
|
Value ct_h = b.create<AtenLinearOp>(hTy, H_prev, weights.R_c, weights.Rb_c);
|
|
|
|
Value ct = b.create<AtenAddTensorOp>(hTy, ct_x, ct_h, cstOne);
|
|
|
|
Value ct_act = createActivationByName(b, activations.g, ct);
|
|
|
|
|
|
|
|
Value C_forget = b.create<AtenMulTensorOp>(hTy, f_act, C_prev);
|
|
|
|
Value C_input = b.create<AtenMulTensorOp>(hTy, i_act, ct_act);
|
|
|
|
|
|
|
|
LstmCellState newCellState;
|
|
|
|
newCellState.C = b.create<AtenAddTensorOp>(hTy, C_forget, C_input, cstOne);
|
|
|
|
Value C_new_act = createActivationByName(b, activations.h, newCellState.C);
|
|
|
|
newCellState.H = b.create<AtenMulTensorOp>(hTy, o_act, C_new_act);
|
|
|
|
return newCellState;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct LstmLayerOutput {
|
|
|
|
Value Y;
|
|
|
|
Value Y_h;
|
|
|
|
Value Y_c;
|
|
|
|
};
|
|
|
|
|
|
|
|
// @brief This function implements the LSTM (Long Short-Term Memory) layer
|
|
|
|
// operation.
|
|
|
|
//
|
|
|
|
// The core computation is performed in a loop that iterates over the sequence
|
|
|
|
// length. In each iteration, it selects the corresponding input, computes the
|
|
|
|
// new hidden state and cell state using the lstm_cell function, and updates the
|
|
|
|
// output tensor.
|
|
|
|
//
|
|
|
|
// @return A struct containing the hidden state history, final hidden state,
|
|
|
|
// and final cell state.
|
|
|
|
LstmLayerOutput lstm_layer(ImplicitLocOpBuilder &b, Value X, Value initial_h,
|
|
|
|
Value initial_c, LstmWeights weights,
|
|
|
|
LstmActivations activations) {
|
|
|
|
|
|
|
|
Location loc = b.getLoc();
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto hTy = cast<ValueTensorType>(initial_h.getType());
|
|
|
|
// these names are snake_case for consistency with onnx.LSTM documentation
|
|
|
|
int64_t seq_len = xTy.getSizes()[0];
|
|
|
|
int64_t batch_size = xTy.getSizes()[1];
|
|
|
|
int64_t input_size = xTy.getSizes()[2];
|
|
|
|
int64_t hidden_size = hTy.getSizes()[1];
|
|
|
|
|
|
|
|
auto cTy = hTy;
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
Value cstSeqLen =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(seq_len));
|
|
|
|
Value cstBatchSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size));
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
|
|
|
|
auto yTy = b.getType<ValueTensorType>(
|
|
|
|
SmallVector<int64_t>{seq_len, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
|
|
|
|
auto YShapeList = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType),
|
|
|
|
ValueRange({cstSeqLen, cstBatchSize, cstHiddenSize}));
|
|
|
|
|
|
|
|
int64_t hDtypeInt =
|
|
|
|
static_cast<int64_t>(getScalarTypeForType(hTy.getDtype()));
|
|
|
|
Value hDtypeIntVal =
|
|
|
|
b.create<ConstantIntOp>(loc, b.getI64IntegerAttr(hDtypeInt));
|
|
|
|
|
|
|
|
Value Y_initial = b.create<AtenZerosOp>(yTy, YShapeList, hDtypeIntVal,
|
|
|
|
cstNone, cstNone, cstNone);
|
|
|
|
|
|
|
|
// Create a for-like PrimLoopOp.
|
|
|
|
Value maxTripCount =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(seq_len));
|
|
|
|
Value loopConditionTrue = b.create<ConstantBoolOp>(true);
|
|
|
|
|
|
|
|
Type loopIndexType = intType;
|
|
|
|
auto loop = b.create<PrimLoopOp>(
|
|
|
|
TypeRange({yTy, hTy, cTy}), maxTripCount, loopConditionTrue,
|
|
|
|
ValueRange({Y_initial, initial_h, initial_c}));
|
|
|
|
{
|
|
|
|
OpBuilder::InsertionGuard guard(b);
|
|
|
|
Block *loopBody =
|
|
|
|
b.createBlock(&loop.getRegion(), loop.getRegion().begin(),
|
|
|
|
TypeRange({
|
|
|
|
loopIndexType,
|
|
|
|
yTy,
|
|
|
|
hTy,
|
|
|
|
cTy,
|
|
|
|
}),
|
|
|
|
{loc, loc, loc, loc} // locs for the loop body arguments
|
|
|
|
);
|
|
|
|
|
|
|
|
Value loopIndex = loopBody->getArgument(0);
|
|
|
|
Value Y_prev = loopBody->getArgument(1);
|
|
|
|
Value H_prev = loopBody->getArgument(2);
|
|
|
|
Value C_prev = loopBody->getArgument(3);
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto XtType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{batch_size, input_size}, xTy.getDtype());
|
|
|
|
|
|
|
|
Value Xt = b.create<AtenSelectIntOp>(XtType, X, cstZero, loopIndex);
|
|
|
|
|
|
|
|
auto [H_new, C_new] =
|
|
|
|
lstm_cell(b, Xt, H_prev, C_prev, weights, activations);
|
|
|
|
|
|
|
|
Type hTyUnsqueezed = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{1, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
Value H_new_unsqueezed =
|
|
|
|
b.create<AtenUnsqueezeOp>(hTyUnsqueezed, H_new, cstZero);
|
|
|
|
|
|
|
|
auto loopIndexPlusOne = b.create<AtenAddIntOp>(intType, loopIndex, cstOne);
|
|
|
|
Value Y_new =
|
|
|
|
b.create<AtenSliceScatterOp>(yTy, Y_prev, H_new_unsqueezed, cstZero,
|
|
|
|
loopIndex, loopIndexPlusOne, cstOne);
|
|
|
|
|
|
|
|
b.create<PrimLoopConditionOp>(loopConditionTrue,
|
|
|
|
ValueRange({Y_new, H_new, C_new}));
|
|
|
|
}
|
|
|
|
LstmLayerOutput output;
|
|
|
|
output.Y = loop.getResult(0);
|
|
|
|
output.Y_h = loop.getResult(1);
|
|
|
|
output.Y_c = loop.getResult(2);
|
|
|
|
return output;
|
|
|
|
}
|
|
|
|
// @brief Expands an ONNX LSTM operation into torch ops.
|
|
|
|
//
|
|
|
|
// This function primarily handles the binding of operands and slicing of the
|
|
|
|
// weight matrix. The majority of the lowering process is managed in the
|
|
|
|
// lstm_layer and lstm_cell. For the shapes and meanings of the inputs, refer to
|
|
|
|
// the ONNX LSTM documentation at:
|
|
|
|
// https://onnx.ai/onnx/operators/onnx__LSTM.html
|
|
|
|
// The variable names are also consistent with the aforementioned documentation.
|
|
|
|
//
|
|
|
|
// This is not e2e tested here but is verified to work numerically downstream in
|
|
|
|
// SHARK-TestSuite.
|
|
|
|
//
|
|
|
|
// TODO: include this test case when the test infrastructure stops initializing
|
|
|
|
// weights separately for the reference and tested layers.
|
|
|
|
// @code{.py}
|
|
|
|
// class LSTMModule(torch.nn.Module):
|
|
|
|
// def __init__(self):
|
|
|
|
// super().__init__()
|
|
|
|
// self.lstm = torch.nn.LSTM(10, 20, 1)
|
|
|
|
// @export
|
|
|
|
// @annotate_args([
|
|
|
|
// None,
|
|
|
|
// ([5, 1, 10], torch.float32, True),
|
|
|
|
// ([1, 1, 20], torch.float32, True),
|
|
|
|
// ([1, 1, 20], torch.float32, True),
|
|
|
|
// ])
|
|
|
|
// def forward(self, input, h0, c0):
|
|
|
|
// return self.lstm(input, (h0, c0))
|
|
|
|
//
|
|
|
|
// @register_test_case(module_factory=LSTMModule)
|
|
|
|
// def LSTMModule_basic(module, tu: TestUtils):
|
|
|
|
// inputs = torch.zeros(5,1,10)
|
|
|
|
// h0 = torch.zeros(1,1,20)
|
|
|
|
// c0 = torch.zeros(1,1,20)
|
|
|
|
//
|
|
|
|
// output, (hn, cn) = module.forward(inputs, h0, c0)
|
|
|
|
// @endcode
|
|
|
|
//
|
|
|
|
// @param binder The OpBinder object used for binding operands.
|
|
|
|
LogicalResult OnnxLstmExpander(OpBinder binder,
|
|
|
|
ConversionPatternRewriter &rewriter) {
|
|
|
|
Location loc = binder.getLoc();
|
|
|
|
mlir::ImplicitLocOpBuilder b(loc, rewriter);
|
|
|
|
|
|
|
|
std::string direction;
|
|
|
|
|
|
|
|
ValueTensorType yTy, Y_hType, Y_cType;
|
|
|
|
if (binder.tensorResultTypeAtIndex(yTy, 0) ||
|
|
|
|
binder.tensorResultTypeAtIndex(Y_hType, 1) ||
|
|
|
|
binder.tensorResultTypeAtIndex(Y_cType, 2)) {
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"At least one outputs must be present");
|
|
|
|
}
|
|
|
|
Value X;
|
|
|
|
if (binder.tensorOperandAtIndex(X, 0))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor X");
|
|
|
|
Value W;
|
|
|
|
if (binder.tensorOperandAtIndex(W, 1))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor W");
|
|
|
|
Value R;
|
|
|
|
if (binder.tensorOperandAtIndex(R, 2))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor R");
|
|
|
|
int64_t hidden_size;
|
|
|
|
if (binder.s64IntegerAttr(hidden_size, "hidden_size"))
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Missing required attribute hidden_size");
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto wTy = cast<ValueTensorType>(W.getType());
|
|
|
|
Value B;
|
|
|
|
if (binder.tensorOperandAtIndex(B, 3)) {
|
|
|
|
B = b.create<AtenZerosOp>(W.getType(), W);
|
|
|
|
}
|
|
|
|
|
|
|
|
llvm::SmallVector<std::string> activationsList;
|
|
|
|
if (binder.stringArrayAttr(activationsList, "activations"))
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Missing required attribute; activations");
|
|
|
|
|
|
|
|
LstmActivations activations;
|
|
|
|
activations.f = "Sigmoid";
|
|
|
|
activations.g = "Tanh";
|
|
|
|
activations.h = "Tanh";
|
|
|
|
if (activationsList.size() == 3) {
|
|
|
|
activations.f = activationsList[0];
|
|
|
|
activations.g = activationsList[1];
|
|
|
|
activations.h = activationsList[2];
|
|
|
|
} else if (activationsList.size() != 0) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "activations must be empty have 3 elements, but " +
|
|
|
|
std::to_string(activationsList.size()) +
|
|
|
|
" are provided.");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!binder.customOpNameStringAttr(direction, "direction", "forward") &&
|
|
|
|
direction != "forward")
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported direction attribute value. "
|
|
|
|
"Only 'forward' is supported but '" +
|
|
|
|
direction + "' is provided.");
|
|
|
|
int64_t num_directions = 1 + (direction == "bidirectional");
|
|
|
|
|
|
|
|
auto XShape = xTy.getSizes();
|
|
|
|
int64_t batch_size = XShape[1];
|
|
|
|
int64_t input_size = XShape[2];
|
|
|
|
if (num_directions != wTy.getSizes()[0])
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "num_directions (" + std::to_string(num_directions) +
|
|
|
|
") does not match the first dimension of wTy (" +
|
|
|
|
std::to_string(wTy.getSizes()[0]) + ")");
|
|
|
|
if (num_directions != 1)
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "num_directions (" + std::to_string(num_directions) +
|
|
|
|
") is not equal to 1");
|
|
|
|
if (4 * hidden_size != wTy.getSizes()[1])
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "4 times hidden_size (" + std::to_string(4 * hidden_size) +
|
|
|
|
") does not match the second dimension of wTy (" +
|
|
|
|
std::to_string(wTy.getSizes()[1]) + ")");
|
|
|
|
if (wTy.getSizes()[2] != input_size)
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op,
|
|
|
|
"The third dimension of wTy (" + std::to_string(wTy.getSizes()[2]) +
|
|
|
|
") does not match input_size (" + std::to_string(input_size) + ")");
|
|
|
|
|
|
|
|
Value W_forward = getDirection(b, 0, W);
|
|
|
|
Value R_forward = getDirection(b, 0, R);
|
|
|
|
Value B_forward = getDirection(b, 0, B);
|
|
|
|
|
|
|
|
auto hTy = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{num_directions, batch_size, hidden_size},
|
|
|
|
xTy.getDtype());
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
|
|
|
|
Value cstNumDirections =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions));
|
|
|
|
Value cstBatchSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size));
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
|
|
|
|
Value hShape = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType),
|
|
|
|
ValueRange({cstNumDirections, cstBatchSize, cstHiddenSize}));
|
|
|
|
|
|
|
|
Value cstDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
|
|
|
|
|
|
|
|
Value initial_h;
|
|
|
|
if (binder.tensorOperandAtIndex(initial_h, 5)) {
|
|
|
|
initial_h =
|
|
|
|
b.create<AtenZerosOp>(hTy, hShape, cstDtype, cstNone, cstNone, cstNone);
|
|
|
|
}
|
|
|
|
Value initial_c;
|
|
|
|
if (binder.tensorOperandAtIndex(initial_c, 6)) {
|
|
|
|
initial_c =
|
|
|
|
b.create<AtenZerosOp>(hTy, hShape, cstDtype, cstNone, cstNone, cstNone);
|
|
|
|
}
|
|
|
|
|
|
|
|
Value initial_h_forward = getDirection(b, 0, initial_h);
|
|
|
|
Value initial_c_forward = getDirection(b, 0, initial_c);
|
|
|
|
|
|
|
|
if (num_directions != 1) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Unsupported num_directions. Only 1 is supported but " +
|
|
|
|
std::to_string(num_directions) + " is provided.");
|
|
|
|
// TODO: support bidirectional LSTM by doing both directions and replacing
|
|
|
|
// Unsqueeze with Stack
|
|
|
|
}
|
|
|
|
// Everything hereon is for the forward direction, with the direction
|
|
|
|
// dimention squeezed out.
|
|
|
|
|
|
|
|
LstmWeights weights; // weights and biases
|
|
|
|
|
|
|
|
auto intConst = [&](int64_t val) {
|
|
|
|
return b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(val));
|
|
|
|
};
|
|
|
|
|
|
|
|
// split B into Wb and Rb
|
|
|
|
Value inputWeightsEndIdx = intConst(4 * hidden_size);
|
|
|
|
Value recurrentWeightsStartIdx = inputWeightsEndIdx;
|
|
|
|
Value recurrentWeightsEndIdx = intConst(8 * hidden_size);
|
|
|
|
auto biasType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size * 4}, wTy.getDtype());
|
|
|
|
Value Wb = b.create<AtenSliceTensorOp>(biasType,
|
|
|
|
/*input=*/B_forward,
|
|
|
|
/*dim=*/cstZero,
|
|
|
|
/*start=*/cstZero,
|
|
|
|
/*end=*/inputWeightsEndIdx,
|
|
|
|
/*step=*/cstOne);
|
|
|
|
Value Rb = b.create<AtenSliceTensorOp>(biasType,
|
|
|
|
/*input=*/B_forward,
|
|
|
|
/*dim=*/cstZero,
|
|
|
|
/*start=*/recurrentWeightsStartIdx,
|
|
|
|
/*end=*/recurrentWeightsEndIdx,
|
|
|
|
/*step=*/cstOne);
|
|
|
|
|
|
|
|
// gate splitting
|
|
|
|
auto gateBiasType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size},
|
|
|
|
cast<ValueTensorType>(Wb.getType()).getDtype());
|
|
|
|
auto gateWeightsTypeIH = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size, input_size},
|
|
|
|
cast<ValueTensorType>(W_forward.getType()).getDtype());
|
|
|
|
auto gateWeightsTypeHH = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size, hidden_size},
|
|
|
|
cast<ValueTensorType>(R_forward.getType()).getDtype());
|
|
|
|
|
|
|
|
Value inputGateWeightsEndIdx = intConst(hidden_size);
|
|
|
|
Value outputGateWeightsEndIdx = intConst(2 * hidden_size);
|
|
|
|
Value forgetGateWeightsEndIdx = intConst(3 * hidden_size);
|
|
|
|
Value cellGateWeightsEndIdx = intConst(4 * hidden_size);
|
|
|
|
|
|
|
|
auto sliceIOFC = [&](std::function<Value(Value, Value)> slicerFunction) {
|
|
|
|
// slice into 4 components and return tuple
|
|
|
|
return std::make_tuple(
|
|
|
|
slicerFunction(cstZero, inputGateWeightsEndIdx),
|
|
|
|
slicerFunction(inputGateWeightsEndIdx, outputGateWeightsEndIdx),
|
|
|
|
slicerFunction(outputGateWeightsEndIdx, forgetGateWeightsEndIdx),
|
|
|
|
slicerFunction(forgetGateWeightsEndIdx, cellGateWeightsEndIdx));
|
|
|
|
};
|
|
|
|
|
|
|
|
auto sliceGateBias = [&](Value startIdx, Value endIdx) {
|
|
|
|
return b.create<AtenSliceTensorOp>(gateBiasType, Wb, cstZero, startIdx,
|
|
|
|
endIdx, cstOne);
|
|
|
|
};
|
|
|
|
std::tie(weights.Wb_i, weights.Wb_o, weights.Wb_f, weights.Wb_c) =
|
|
|
|
sliceIOFC(sliceGateBias);
|
|
|
|
|
|
|
|
auto sliceGateBiasR = [&](Value startIdx, Value endIdx) {
|
|
|
|
return b.create<AtenSliceTensorOp>(gateBiasType, Rb, cstZero, startIdx,
|
|
|
|
endIdx, cstOne);
|
|
|
|
};
|
|
|
|
std::tie(weights.Rb_i, weights.Rb_o, weights.Rb_f, weights.Rb_c) =
|
|
|
|
sliceIOFC(sliceGateBiasR);
|
|
|
|
|
|
|
|
auto sliceGateWeightsIH = [&](Value startIdx, Value endIdx) {
|
|
|
|
return b.create<AtenSliceTensorOp>(gateWeightsTypeIH, W_forward, cstZero,
|
|
|
|
startIdx, endIdx, cstOne);
|
|
|
|
};
|
|
|
|
std::tie(weights.W_i, weights.W_o, weights.W_f, weights.W_c) =
|
|
|
|
sliceIOFC(sliceGateWeightsIH);
|
|
|
|
|
|
|
|
auto sliceGateWeightsHH = [&](Value startIdx, Value endIdx) {
|
|
|
|
return b.create<AtenSliceTensorOp>(gateWeightsTypeHH, R_forward, cstZero,
|
|
|
|
startIdx, endIdx, cstOne);
|
|
|
|
};
|
|
|
|
std::tie(weights.R_i, weights.R_o, weights.R_f, weights.R_c) =
|
|
|
|
sliceIOFC(sliceGateWeightsHH);
|
|
|
|
LstmLayerOutput lstmLayerOutput = lstm_layer(
|
|
|
|
b, X, initial_h_forward, initial_c_forward, weights, activations);
|
|
|
|
|
|
|
|
auto Y_h_Y_c_unsqueezed_type = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{num_directions, batch_size, hidden_size},
|
|
|
|
cast<ValueTensorType>(lstmLayerOutput.Y_h.getType()).getDtype());
|
|
|
|
Value Y_h_unsqueezed = b.create<AtenUnsqueezeOp>(
|
|
|
|
Y_h_Y_c_unsqueezed_type, lstmLayerOutput.Y_h, cstZero);
|
|
|
|
Value Y_c_unsqueezed = b.create<AtenUnsqueezeOp>(
|
|
|
|
Y_h_Y_c_unsqueezed_type, lstmLayerOutput.Y_c, cstZero);
|
|
|
|
|
|
|
|
// unsqueeze num_directions dim1 of Y
|
|
|
|
// to create the onnx.LSTM output shape [seq_length, num_directions,
|
|
|
|
// batch_size, hidden_size]
|
|
|
|
Value Y_unsqueezed =
|
|
|
|
b.create<AtenUnsqueezeOp>(yTy, lstmLayerOutput.Y, cstOne);
|
|
|
|
|
|
|
|
rewriter.replaceOp(binder.op, mlir::ValueRange{Y_unsqueezed, Y_h_unsqueezed,
|
|
|
|
Y_c_unsqueezed});
|
|
|
|
return success();
|
|
|
|
}
|
|
|
|
|
|
|
|
// W[zrh] - W parameter weight matrix for update, reset, and hidden gates
|
|
|
|
// R[zrh] - R recurrence weight matrix for update, reset, and hidden gates
|
|
|
|
// Wb[zrh] - W bias vectors for update, reset, and hidden gates
|
|
|
|
// Rb[zrh] - R bias vectors for update, reset, and hidden gates
|
|
|
|
// backwards currently not supported
|
|
|
|
|
|
|
|
struct GruWeights {
|
|
|
|
Value Wz;
|
|
|
|
Value Wr;
|
|
|
|
Value Wh;
|
|
|
|
Value Rz;
|
|
|
|
Value Rr;
|
|
|
|
Value Rh;
|
|
|
|
Value Wbz;
|
|
|
|
Value Wbr;
|
|
|
|
Value Wbh;
|
|
|
|
Value Rbz;
|
|
|
|
Value Rbr;
|
|
|
|
Value Rbh;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct GruLayerOutput {
|
|
|
|
Value Y;
|
|
|
|
Value Y_h;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct GruActivations {
|
|
|
|
std::string f;
|
|
|
|
std::string g;
|
|
|
|
};
|
|
|
|
|
|
|
|
Value gru_cell(ImplicitLocOpBuilder &b, Value Xt, Value H_prev,
|
|
|
|
GruWeights weights, GruActivations activations,
|
|
|
|
bool linear_before_reset) {
|
|
|
|
auto hTy = cast<ValueTensorType>(H_prev.getType());
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
|
|
|
|
Value z_w = b.create<AtenLinearOp>(hTy, Xt, weights.Wz, weights.Wbz);
|
|
|
|
Value z_r = b.create<AtenLinearOp>(hTy, H_prev, weights.Rz, weights.Rbz);
|
|
|
|
Value z_pre = b.create<AtenAddTensorOp>(hTy, z_w, z_r, cstOne);
|
|
|
|
Value zt = createActivationByName(b, activations.f, z_pre);
|
|
|
|
|
|
|
|
Value r_w = b.create<AtenLinearOp>(hTy, Xt, weights.Wr, weights.Wbr);
|
|
|
|
Value r_r = b.create<AtenLinearOp>(hTy, H_prev, weights.Rr, weights.Rbr);
|
|
|
|
Value r_pre = b.create<AtenAddTensorOp>(hTy, r_w, r_r, cstOne);
|
|
|
|
Value rt = createActivationByName(b, activations.f, r_pre);
|
|
|
|
|
|
|
|
Value h_w = b.create<AtenLinearOp>(hTy, Xt, weights.Wh, weights.Wbh);
|
|
|
|
Value h_r;
|
|
|
|
if (linear_before_reset) {
|
|
|
|
// when linear_before_reset = 1, multiply r with H_prev to reset
|
|
|
|
// before applying linear layer
|
|
|
|
Value h_linear =
|
|
|
|
b.create<AtenLinearOp>(hTy, H_prev, weights.Rh, weights.Rbh);
|
|
|
|
h_r = b.create<AtenMulTensorOp>(hTy, h_linear, rt);
|
|
|
|
} else {
|
|
|
|
// otherwise, multiply first and then apply linear layer
|
|
|
|
Value h_reset = b.create<AtenMulTensorOp>(hTy, H_prev, rt);
|
|
|
|
h_r = b.create<AtenLinearOp>(hTy, h_reset, weights.Rh, weights.Rbh);
|
|
|
|
}
|
|
|
|
Value h_pre = b.create<AtenAddTensorOp>(hTy, h_w, h_r, cstOne);
|
|
|
|
Value ht = createActivationByName(b, activations.g, h_pre);
|
|
|
|
|
|
|
|
// Create a constant tensor filled with ones, matching the shape of zt
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
int64_t typeInt = (int64_t)getScalarTypeForType(hTy.getDtype());
|
|
|
|
Value dtype = b.create<ConstantIntOp>(b.getI64IntegerAttr(typeInt));
|
|
|
|
Value ones = b.create<Torch::AtenOnesLikeOp>(
|
|
|
|
hTy, zt, dtype, /*layout=*/cstNone,
|
|
|
|
/*device=*/cstNone, /*pin_memory=*/cstNone, /*memory_format=*/cstNone);
|
|
|
|
|
|
|
|
Value one_minus_zt = b.create<AtenSubTensorOp>(hTy, ones, zt, cstOne);
|
|
|
|
Value ht_scaled = b.create<AtenMulTensorOp>(hTy, one_minus_zt, ht);
|
|
|
|
Value H_prev_zt = b.create<AtenMulTensorOp>(hTy, H_prev, zt);
|
|
|
|
Value H_new = b.create<AtenAddTensorOp>(hTy, ht_scaled, H_prev_zt, cstOne);
|
|
|
|
|
|
|
|
return H_new;
|
|
|
|
}
|
|
|
|
|
|
|
|
GruLayerOutput gru_layer(ImplicitLocOpBuilder &b, Value X, Value initial_h,
|
|
|
|
GruWeights weights, GruActivations activations,
|
|
|
|
bool linear_before_reset) {
|
|
|
|
Location loc = b.getLoc();
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto hTy = cast<ValueTensorType>(initial_h.getType());
|
|
|
|
|
|
|
|
// Get sizes and store them in intermediate variables
|
|
|
|
auto xTySizes = xTy.getSizes();
|
|
|
|
auto hTySizes = hTy.getSizes();
|
|
|
|
|
|
|
|
int64_t seq_len = xTySizes[0];
|
|
|
|
int64_t batch_size = xTySizes[1];
|
|
|
|
int64_t input_size = xTySizes[2];
|
|
|
|
int64_t hidden_size = hTySizes[1];
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
Value cstSeqLen =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(seq_len));
|
|
|
|
Value cstBatchSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size));
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
|
|
|
|
auto yTy = b.getType<ValueTensorType>(
|
|
|
|
SmallVector<int64_t>{seq_len, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
|
|
|
|
auto YShapeList = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType),
|
|
|
|
ValueRange({cstSeqLen, cstBatchSize, cstHiddenSize}));
|
|
|
|
|
|
|
|
int64_t hDtypeInt =
|
|
|
|
static_cast<int64_t>(getScalarTypeForType(hTy.getDtype()));
|
|
|
|
Value hDtypeIntVal = b.create<ConstantIntOp>(b.getI64IntegerAttr(hDtypeInt));
|
|
|
|
|
|
|
|
Value Y_initial = b.create<AtenZerosOp>(yTy, YShapeList, hDtypeIntVal,
|
|
|
|
cstNone, cstNone, cstNone);
|
|
|
|
|
|
|
|
Value maxTripCount = cstSeqLen;
|
|
|
|
Value loopConditionTrue = b.create<ConstantBoolOp>(true);
|
|
|
|
|
|
|
|
Type loopIndexType = intType;
|
|
|
|
|
|
|
|
auto loop = b.create<PrimLoopOp>(TypeRange({yTy, hTy}), maxTripCount,
|
|
|
|
loopConditionTrue,
|
|
|
|
ValueRange({Y_initial, initial_h}));
|
|
|
|
|
|
|
|
{
|
|
|
|
OpBuilder::InsertionGuard guard(b);
|
|
|
|
Block *loopBody =
|
|
|
|
b.createBlock(&loop.getRegion(), loop.getRegion().begin(),
|
|
|
|
TypeRange({loopIndexType, yTy, hTy}), {loc, loc, loc});
|
|
|
|
|
|
|
|
Value loopIndex = loopBody->getArgument(0);
|
|
|
|
Value Y_prev = loopBody->getArgument(1);
|
|
|
|
Value H_prev = loopBody->getArgument(2);
|
|
|
|
|
|
|
|
auto XtType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{batch_size, input_size}, xTy.getDtype());
|
|
|
|
|
|
|
|
Value Xt = b.create<AtenSelectIntOp>(XtType, X, cstZero, loopIndex);
|
|
|
|
|
|
|
|
Value H_new =
|
|
|
|
gru_cell(b, Xt, H_prev, weights, activations, linear_before_reset);
|
|
|
|
|
|
|
|
Type hTyUnsqueezed = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{1, batch_size, hidden_size}, hTy.getDtype());
|
|
|
|
Value H_new_unsqueezed =
|
|
|
|
b.create<AtenUnsqueezeOp>(hTyUnsqueezed, H_new, cstZero);
|
|
|
|
|
|
|
|
auto loopIndexPlusOne = b.create<AtenAddIntOp>(intType, loopIndex, cstOne);
|
|
|
|
Value Y_new =
|
|
|
|
b.create<AtenSliceScatterOp>(yTy, Y_prev, H_new_unsqueezed, cstZero,
|
|
|
|
loopIndex, loopIndexPlusOne, cstOne);
|
|
|
|
|
|
|
|
b.create<PrimLoopConditionOp>(loopConditionTrue,
|
|
|
|
ValueRange({Y_new, H_new}));
|
|
|
|
}
|
|
|
|
|
|
|
|
GruLayerOutput output;
|
|
|
|
output.Y = loop.getResult(0);
|
|
|
|
output.Y_h = loop.getResult(1);
|
|
|
|
|
|
|
|
return output;
|
|
|
|
}
|
|
|
|
|
|
|
|
LogicalResult OnnxGruExpander(OpBinder binder,
|
|
|
|
ConversionPatternRewriter &rewriter) {
|
|
|
|
Location loc = binder.getLoc();
|
|
|
|
mlir::ImplicitLocOpBuilder b(loc, rewriter);
|
|
|
|
|
|
|
|
auto intType = b.getType<IntType>();
|
|
|
|
Value cstNone = b.create<ConstantNoneOp>();
|
|
|
|
Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
|
|
|
|
Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
|
|
|
|
Value cstTwo = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2));
|
|
|
|
|
|
|
|
// Binding arguments
|
|
|
|
ValueTensorType yTy, Y_hType;
|
|
|
|
if (binder.tensorResultTypeAtIndex(yTy, 0) ||
|
|
|
|
binder.tensorResultTypeAtIndex(Y_hType, 1)) {
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"At least one output must be present");
|
|
|
|
}
|
|
|
|
|
|
|
|
Value X, W, R, B, initial_h, sequence_lens;
|
|
|
|
if (binder.tensorOperandAtIndex(X, 0) || binder.tensorOperandAtIndex(W, 1) ||
|
|
|
|
binder.tensorOperandAtIndex(R, 2))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Missing required input tensor");
|
|
|
|
|
|
|
|
if (binder.tensorOperandAtIndex(B, 3)) {
|
|
|
|
// if no b found, set to null and create one later
|
|
|
|
B = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
int64_t hidden_size;
|
|
|
|
if (binder.s64IntegerAttr(hidden_size, "hidden_size"))
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Missing required attribute hidden_size");
|
|
|
|
|
|
|
|
auto xTy = cast<ValueTensorType>(X.getType());
|
|
|
|
auto wTy = cast<ValueTensorType>(W.getType());
|
|
|
|
|
|
|
|
// Setting up activations
|
|
|
|
GruActivations activations;
|
|
|
|
activations.f = "Sigmoid";
|
|
|
|
activations.g = "Tanh";
|
|
|
|
|
|
|
|
llvm::SmallVector<std::string> activationsList;
|
|
|
|
if (!binder.stringArrayAttr(activationsList, "activations") &&
|
|
|
|
activationsList.size() == 2) {
|
|
|
|
activations.f = activationsList[0];
|
|
|
|
activations.g = activationsList[1];
|
|
|
|
} else if (activationsList.size() > 0) {
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Unsupported number of activation functions");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Other attributes
|
|
|
|
int64_t layout;
|
|
|
|
if (binder.s64IntegerAttr(layout, "layout", 0))
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported layout attribute type.");
|
|
|
|
|
|
|
|
std::string direction;
|
|
|
|
if (!binder.customOpNameStringAttr(direction, "direction", "forward") &&
|
|
|
|
direction != "forward")
|
|
|
|
return rewriter.notifyMatchFailure(binder.op,
|
|
|
|
"Unsupported direction attribute value");
|
|
|
|
|
|
|
|
int64_t num_directions = direction == "bidirectional" ? 2 : 1;
|
|
|
|
// Validations
|
|
|
|
auto XShape = xTy.getSizes();
|
|
|
|
int64_t batch_size = (layout == 0) ? XShape[1] : XShape[0];
|
|
|
|
int64_t input_size = XShape[2];
|
|
|
|
|
|
|
|
std::ostringstream oss;
|
|
|
|
|
|
|
|
if (num_directions != 1) {
|
|
|
|
oss << "Expected num_directions to be 1, but got " << num_directions
|
|
|
|
<< ". ";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (hidden_size * 3 != wTy.getSizes()[1]) {
|
|
|
|
oss << "Expected dim 1 of W to be the same as 3*hidden_size "
|
|
|
|
<< 3 * hidden_size << ", but got " << wTy.getSizes()[1] << ". ";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (wTy.getSizes()[2] != input_size) {
|
|
|
|
oss << "Expected wTy.getSizes()[2] to be " << input_size << ", but got "
|
|
|
|
<< wTy.getSizes()[2] << ". ";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!oss.str().empty()) {
|
|
|
|
return rewriter.notifyMatchFailure(binder.op, oss.str());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Setting up initial_h
|
|
|
|
auto hTy = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{num_directions, batch_size, hidden_size},
|
|
|
|
xTy.getDtype());
|
|
|
|
|
|
|
|
if (binder.tensorOperandAtIndex(initial_h, 5)) {
|
|
|
|
Value cstNumDirections =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions));
|
|
|
|
Value cstBatchSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(batch_size));
|
|
|
|
Value cstHiddenSize =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(hidden_size));
|
|
|
|
Value hShape = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType),
|
|
|
|
ValueRange({cstNumDirections, cstBatchSize, cstHiddenSize}));
|
|
|
|
Value cstDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
|
|
|
|
initial_h =
|
|
|
|
b.create<AtenZerosOp>(hTy, hShape, cstDtype, cstNone, cstNone, cstNone);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (binder.tensorOperandAtIndex(sequence_lens, 4))
|
|
|
|
sequence_lens = b.create<ConstantNoneOp>();
|
|
|
|
|
|
|
|
float clip;
|
|
|
|
if (!binder.f32FloatAttr(clip, "clip") && clip != 0.0f)
|
|
|
|
return rewriter.notifyMatchFailure(
|
|
|
|
binder.op, "Clip not supported (specified with a value of " +
|
|
|
|
std::to_string(clip) + ")");
|
|
|
|
|
|
|
|
int64_t linear_before_reset_int;
|
|
|
|
if (binder.s64IntegerAttr(linear_before_reset_int, "linear_before_reset", 0))
|
|
|
|
linear_before_reset_int = 0;
|
|
|
|
bool linear_before_reset = linear_before_reset_int != 0;
|
|
|
|
|
|
|
|
// fill in B
|
|
|
|
Value cstXDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
|
|
|
|
if (B == nullptr) {
|
|
|
|
SmallVector<int64_t> BShape = {num_directions, 2 * hidden_size};
|
|
|
|
SmallVector<Value> BShapeListContents = {
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions)),
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2 * hidden_size))};
|
|
|
|
Value BShapeList = b.create<PrimListConstructOp>(
|
|
|
|
b.getType<ListType>(intType), BShapeListContents);
|
|
|
|
auto BType = b.getType<ValueTensorType>(BShape, wTy.getDtype());
|
|
|
|
B = b.create<Torch::AtenZerosOp>(BType, BShapeList, cstXDtype, cstNone,
|
|
|
|
cstNone, cstNone);
|
|
|
|
}
|
|
|
|
|
|
|
|
Value W_forward = getDirection(b, 0, W);
|
|
|
|
Value R_forward = getDirection(b, 0, R);
|
|
|
|
Value B_forward = getDirection(b, 0, B);
|
|
|
|
Value initial_h_forward = getDirection(b, 0, initial_h);
|
|
|
|
|
|
|
|
GruWeights weights;
|
|
|
|
|
|
|
|
// Slice a tensor into numSlices slices of size sliceSize
|
|
|
|
// This is used for slicing the weights & biases into the individual gates
|
|
|
|
auto sliceTensor = [&](Value tensor, int64_t sliceSize, int64_t numSlices,
|
|
|
|
ValueTensorType sliceType) {
|
|
|
|
SmallVector<Value> slices;
|
|
|
|
for (int64_t i = 0; i < numSlices; ++i) {
|
|
|
|
Value start =
|
|
|
|
b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(i * sliceSize));
|
|
|
|
Value end = b.create<ConstantIntOp>(
|
|
|
|
intType, b.getI64IntegerAttr((i + 1) * sliceSize));
|
|
|
|
|
|
|
|
Value slice = b.create<AtenSliceTensorOp>(sliceType, tensor,
|
|
|
|
cstZero, // dim to slice on
|
|
|
|
start, end,
|
|
|
|
cstOne // step
|
|
|
|
);
|
|
|
|
|
|
|
|
slices.push_back(slice);
|
|
|
|
}
|
|
|
|
return slices;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Slice W
|
|
|
|
auto wSliceType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size, input_size}, wTy.getDtype());
|
|
|
|
auto W_slices = sliceTensor(W_forward, hidden_size, 3, wSliceType);
|
|
|
|
std::tie(weights.Wz, weights.Wr, weights.Wh) =
|
|
|
|
std::make_tuple(W_slices[0], W_slices[1], W_slices[2]);
|
|
|
|
|
|
|
|
// Slice R
|
|
|
|
auto rSliceType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size, hidden_size}, wTy.getDtype());
|
|
|
|
auto R_slices = sliceTensor(R_forward, hidden_size, 3, rSliceType);
|
|
|
|
std::tie(weights.Rz, weights.Rr, weights.Rh) =
|
|
|
|
std::make_tuple(R_slices[0], R_slices[1], R_slices[2]);
|
|
|
|
|
|
|
|
// Slice B
|
|
|
|
auto bSliceType = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{hidden_size}, wTy.getDtype());
|
|
|
|
auto B_slices = sliceTensor(B_forward, hidden_size, 6, bSliceType);
|
|
|
|
std::tie(weights.Wbz, weights.Wbr, weights.Wbh, weights.Rbz, weights.Rbr,
|
|
|
|
weights.Rbh) =
|
|
|
|
std::make_tuple(B_slices[0], B_slices[1], B_slices[2], B_slices[3],
|
|
|
|
B_slices[4], B_slices[5]);
|
|
|
|
|
|
|
|
// Process inputs based on layout
|
|
|
|
Value X_processed, initial_h_processed;
|
|
|
|
ValueTensorType yTy_processed, Y_hType_processed;
|
|
|
|
|
|
|
|
if (layout == 0) {
|
|
|
|
X_processed = X;
|
|
|
|
initial_h_processed = initial_h_forward;
|
|
|
|
yTy_processed = yTy;
|
|
|
|
Y_hType_processed = Y_hType;
|
|
|
|
} else {
|
|
|
|
X_processed = b.create<AtenTransposeIntOp>(X.getType(), X, cstZero, cstOne);
|
|
|
|
initial_h_processed = b.create<AtenTransposeIntOp>(
|
|
|
|
initial_h.getType(), initial_h_forward, cstZero, cstOne);
|
|
|
|
|
|
|
|
auto yTySizes = yTy.getSizes();
|
|
|
|
auto Y_hTypeSizes = Y_hType.getSizes();
|
|
|
|
|
|
|
|
yTy_processed = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{yTySizes[1], yTySizes[0], yTySizes[2],
|
|
|
|
yTySizes[3]},
|
|
|
|
yTy.getDtype());
|
|
|
|
|
|
|
|
Y_hType_processed = b.getType<ValueTensorType>(
|
|
|
|
llvm::SmallVector<int64_t>{Y_hTypeSizes[1], Y_hTypeSizes[0],
|
|
|
|
Y_hTypeSizes[2]},
|
|
|
|
Y_hType.getDtype());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Weights and biases ready. Calling GRU layer to insert the actual ops.
|
|
|
|
GruLayerOutput gruLayerOutput =
|
|
|
|
gru_layer(b, X_processed, initial_h_processed, weights, activations,
|
|
|
|
linear_before_reset);
|
|
|
|
|
|
|
|
// Process outputs based on layout
|
|
|
|
Value Y_final, Y_h_final;
|
|
|
|
if (layout == 0) {
|
|
|
|
Y_final = b.create<AtenUnsqueezeOp>(yTy, gruLayerOutput.Y, cstOne);
|
|
|
|
Y_h_final = b.create<AtenUnsqueezeOp>(Y_hType, gruLayerOutput.Y_h, cstZero);
|
|
|
|
} else {
|
|
|
|
auto Y_transposed = b.create<AtenTransposeIntOp>(
|
|
|
|
gruLayerOutput.Y.getType(), gruLayerOutput.Y, cstZero, cstOne);
|
|
|
|
Y_final = b.create<AtenUnsqueezeOp>(yTy, Y_transposed, cstTwo);
|
|
|
|
|
|
|
|
auto Y_h_transposed = b.create<AtenTransposeIntOp>(
|
|
|
|
gruLayerOutput.Y_h.getType(), gruLayerOutput.Y_h, cstZero, cstOne);
|
|
|
|
Y_h_final = b.create<AtenUnsqueezeOp>(Y_hType, Y_h_transposed, cstZero);
|
|
|
|
}
|
|
|
|
|
|
|
|
rewriter.replaceOp(binder.op, mlir::ValueRange{Y_final, Y_h_final});
|
|
|
|
return success();
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace mlir::torch::onnx_c
|