Skip to content

[TorchToLinalg] Use Op with native channel order for quantized conv2d #3807

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 22, 2024
Merged

[TorchToLinalg] Use Op with native channel order for quantized conv2d #3807

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
59 changes: 31 additions & 28 deletions lib/Conversion/TorchToLinalg/Linear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1125,54 +1125,57 @@ class ConvertAtenConvolutionOp : public OpConversionPattern<AtenConvolutionOp> {
}

if (numGroups == 1 && inputZp) {
// The quantized version uses a different channel ordering so we need to
// permute the tensors in order to use the existing path. We should
// eventually directly support this channel ordering.
llvm::SmallVector<int64_t> inPerms, weightPerms;
inPerms.push_back(0); // N stays at the front for input.
// Then we expect the spatial dimensions
for (size_t i = 0; i < numSpatialDims; ++i) {
inPerms.push_back(i + 2);
weightPerms.push_back(i + 2);
}
inPerms.push_back(1);
weightPerms.append({1, 0});

paddedInput = transposeValue(op.getLoc(), paddedInput, inPerms, rewriter);
weight = transposeValue(op.getLoc(), weight, weightPerms, rewriter);
outputTensor =
transposeValue(op.getLoc(), outputTensor, inPerms, rewriter);

switch (numSpatialDims) {
case 2:
conv = rewriter
.create<linalg::Conv2DNhwcHwcfQOp>(
.create<linalg::Conv2DNchwFchwQOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight, inputZp, weightZp},
outputTensor, stridesAttr, dilationAttr)
.getResult(0);
break;
case 3:
case 3: {
// The quantized version uses a different channel ordering so we need to
// permute the tensors in order to use the existing path. We should
// eventually directly support this channel ordering.
llvm::SmallVector<int64_t> inPerms, weightPerms;
inPerms.push_back(0); // N stays at the front for input.
// Then we expect the spatial dimensions
for (size_t i = 0; i < numSpatialDims; ++i) {
inPerms.push_back(i + 2);
weightPerms.push_back(i + 2);
}
inPerms.push_back(1);
weightPerms.append({1, 0});

paddedInput =
transposeValue(op.getLoc(), paddedInput, inPerms, rewriter);
weight = transposeValue(op.getLoc(), weight, weightPerms, rewriter);
outputTensor =
transposeValue(op.getLoc(), outputTensor, inPerms, rewriter);

conv = rewriter
.create<linalg::Conv3DNdhwcDhwcfQOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight, inputZp, weightZp},
outputTensor, stridesAttr, dilationAttr)
.getResult(0);

llvm::SmallVector<int64_t> outPerms;
outPerms.push_back(0);
outPerms.push_back(inPerms.size() - 1);
for (size_t i = 0; i < numSpatialDims; ++i) {
outPerms.push_back(i + 1);
}
conv = transposeValue(op.getLoc(), conv, outPerms, rewriter);

break;
}
default:
return rewriter.notifyMatchFailure(
op, "unimplemented: only 1D, 2D, and 3D convolution supported");
};

llvm::SmallVector<int64_t> outPerms;
outPerms.push_back(0);
outPerms.push_back(inPerms.size() - 1);
for (size_t i = 0; i < numSpatialDims; ++i) {
outPerms.push_back(i + 1);
}
conv = transposeValue(op.getLoc(), conv, outPerms, rewriter);

Type newResultType = getTypeConverter()->convertType(op.getType());
if (accumulatorDType != resultDTy) {
Type resultElementType =
Expand Down
8 changes: 2 additions & 6 deletions test/Conversion/TorchToLinalg/convolution.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,12 +24,8 @@ func.func @torch.aten.convolution$nobias(%arg0: !torch.vtensor<[1,24,16,128,128]
// CHECK: %[[c7:.*]] = arith.constant 7 : i32
// CHECK: %[[input:.*]] = torch_c.to_builtin_tensor %arg0 : !torch.vtensor<[?,?,?,?],si8> -> tensor<?x?x?x?xi8>
// CHECK: %[[weight:.*]] = torch_c.to_builtin_tensor %arg1 : !torch.vtensor<[?,?,?,?],si8> -> tensor<?x?x?x?xi8>
// CHECK: %[[TransInput:.*]] = linalg.transpose ins(%[[input]] : tensor<?x?x?x?xi8>)
// CHECK-SAME: permutation = [0, 2, 3, 1]
// CHECK: %[[TransWeight:.*]] = linalg.transpose ins(%[[weight]] : tensor<?x?x?x?xi8>)
// CHECK-SAME: permutation = [2, 3, 1, 0]
// CHECK: %[[conv:.*]] = linalg.conv_2d_nhwc_hwcf_q {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>}
// CHECK-SAME: ins(%[[TransInput]], %[[TransWeight]], %[[c7]], %[[c3]] : tensor<?x?x?x?xi8>, tensor<?x?x?x?xi8>, i32, i32)
// CHECK: %[[conv:.*]] = linalg.conv_2d_nchw_fchw_q {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>}
// CHECK-SAME: ins(%[[input]], %[[weight]], %[[c7]], %[[c3]] : tensor<?x?x?x?xi8>, tensor<?x?x?x?xi8>, i32, i32)
// CHECK-SAME: outs(%[[convout:.*]] : tensor<?x?x?x?xi32>) -> tensor<?x?x?x?xi32>
func.func @q_conv_test(%arg0: !torch.vtensor<[?,?,?,?],si8>, %arg1: !torch.vtensor<[?,?,?,?],si8>, %arg2: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?,?,?,?],f32> {
%false = torch.constant.bool false
Expand Down
Loading