Skip to content

[mlir][linalg] Add Grouped Convolution Ops: conv_2d_nhwgc_gfhwc and conv_2d_nhwgc_gfhwc_q #108192

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
Nov 8, 2024
+330 −0
Merged

[mlir][linalg] Add Grouped Convolution Ops: conv_2d_nhwgc_gfhwc and conv_2d_nhwgc_gfhwc_q #108192

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
237 changes: 237 additions & 0 deletions mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3475,6 +3475,243 @@ structured_op: !LinalgStructuredOpConfig
- !ScalarExpression
scalar_arg: K
--- !LinalgOpConfig
metadata: !LinalgOpMetadata
name: conv_2d_nhwgc_gfhwc
cpp_class_name: Conv2DNhwgcGfhwcOp
doc: |-
Performs 2-D grouped convolution.

Layout:
* Input: NHWGC.
* Kernel: GFHWC.

Numeric casting is performed on the operands to the inner multiply, promoting
them to the same data type as the accumulator/output.
implements:
- LinalgConvolutionOpInterface
structured_op: !LinalgStructuredOpConfig
args:
- !LinalgOperandDefConfig
name: I
kind: input_tensor
type_var: T1
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s0, s1 * s2 + s3 * s4, s5 * s6 + s7 * s8, s9, s10)>
- !LinalgOperandDefConfig
name: K
kind: input_tensor
type_var: T2
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s9, s11, s3, s7, s10)>
- !LinalgOperandDefConfig
name: O
kind: output_tensor
type_var: U
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s0, s1, s5, s9, s11)>
- !LinalgOperandDefConfig
name: strides
kind: index_attr
index_attr_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
-> (s2, s6)>
default_indices:
- 1
- 1
- !LinalgOperandDefConfig
name: dilations
kind: index_attr
index_attr_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
-> (s4, s8)>
default_indices:
- 1
- 1
indexing_maps: !LinalgIndexingMapsConfig
static_indexing_maps:
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d0, d1 * s2 + d5 * s4, d2 * s6 + d6 * s8, d3, d7)>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d3, d4, d5, d6, d7)>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d0, d1, d2, d3, d4)>
iterator_types:
- parallel
- parallel
- parallel
- parallel
- parallel
- reduction
- reduction
- reduction
assignments:
- !ScalarAssign
arg: O
value: !ScalarExpression
scalar_fn:
kind: binary
fn_name: add
operands:
- !ScalarExpression
scalar_arg: O
- !ScalarExpression
scalar_fn:
kind: binary
fn_name: mul
operands:
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: I
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: K
--- !LinalgOpConfig
metadata: !LinalgOpMetadata
name: conv_2d_nhwgc_gfhwc_q
cpp_class_name: Conv2DNhwgcGfhwcQOp
doc: |-
Performs 2-D grouped convolution with zero point offsets.

Layout:
* Input: NHWGC.
* Kernel: GFHWC.

Numeric casting is performed on the operands to the inner multiply, promoting
them to the same data type as the accumulator/output. This includes the zero
point offsets common to quantized operations.
implements:
- LinalgConvolutionOpInterface
structured_op: !LinalgStructuredOpConfig
args:
- !LinalgOperandDefConfig
name: I
kind: input_tensor
type_var: T1
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s0, s1 * s2 + s3 * s4, s5 * s6 + s7 * s8, s9, s10)>
- !LinalgOperandDefConfig
name: K
kind: input_tensor
type_var: T2
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s9, s11, s3, s7, s10)>
- !LinalgOperandDefConfig
name: IZp
kind: scalar
type_var: I32
- !LinalgOperandDefConfig
name: KZp
kind: scalar
type_var: I32
- !LinalgOperandDefConfig
name: O
kind: output_tensor
type_var: U
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
(s0, s1, s5, s9, s11)>
- !LinalgOperandDefConfig
name: strides
kind: index_attr
index_attr_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
-> (s2, s6)>
default_indices:
- 1
- 1
- !LinalgOperandDefConfig
name: dilations
kind: index_attr
index_attr_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
-> (s4, s8)>
default_indices:
- 1
- 1
indexing_maps: !LinalgIndexingMapsConfig
static_indexing_maps:
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d0, d1 * s2 + d5 * s4, d2 * s6 + d6 * s8, d3, d7)>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d3, d4, d5, d6, d7)>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> ()>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> ()>
- affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
s8, s9, s10, s11] -> (d0, d1, d2, d3, d4)>
iterator_types:
- parallel
- parallel
- parallel
- parallel
- parallel
- reduction
- reduction
- reduction
assignments:
- !ScalarAssign
arg: O
value: !ScalarExpression
scalar_fn:
kind: binary
fn_name: add
operands:
- !ScalarExpression
scalar_arg: O
- !ScalarExpression
scalar_fn:
kind: binary
fn_name: mul
operands:
- !ScalarExpression
scalar_fn:
kind: binary
fn_name: sub
operands:
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: I
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: IZp
- !ScalarExpression
scalar_fn:
kind: binary
fn_name: sub
operands:
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: K
- !ScalarExpression
scalar_fn:
kind: type
fn_name: cast_signed
type_var: U
operands:
- !ScalarExpression
scalar_arg: KZp
--- !LinalgOpConfig
metadata: !LinalgOpMetadata
name: conv_2d_ngchw_gfchw_q
cpp_class_name: Conv2DNgchwGfchwQOp
Expand Down
61 changes: 61 additions & 0 deletions mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -964,6 +964,67 @@ def conv_2d_ngchw_gfchw(
) * TypeFn.cast_signed(U, K[D.g, D.fg, D.c, D.kh, D.kw])


@linalg_structured_op
def conv_2d_nhwgc_gfhwc(
I=TensorDef(
T1, S.N, S.OH * S.SH + S.KH * S.DH, S.OW * S.SW + S.KW * S.DW, S.G, S.C
),
K=TensorDef(T2, S.G, S.FG, S.KH, S.KW, S.C),
O=TensorDef(U, S.N, S.OH, S.OW, S.G, S.FG, output=True),
strides=IndexAttrDef(S.SH, S.SW, default=[1, 1]),
dilations=IndexAttrDef(S.DH, S.DW, default=[1, 1]),
):
"""Performs 2-D grouped convolution.

Layout:
* Input: NHWGC.
* Kernel: GFHWC.

Numeric casting is performed on the operands to the inner multiply, promoting
them to the same data type as the accumulator/output.
"""
implements(ConvolutionOpInterface)
domain(D.n, D.oh, D.ow, D.g, D.fg, D.kh, D.kw, D.c)
O[D.n, D.oh, D.ow, D.g, D.fg] += TypeFn.cast_signed(
U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW, D.g, D.c]
) * TypeFn.cast_signed(U, K[D.g, D.fg, D.kh, D.kw, D.c])


@linalg_structured_op
def conv_2d_nhwgc_gfhwc_q(
I=TensorDef(
T1, S.N, S.OH * S.SH + S.KH * S.DH, S.OW * S.SW + S.KW * S.DW, S.G, S.C
),
K=TensorDef(T2, S.G, S.FG, S.KH, S.KW, S.C),
IZp=ScalarDef(I32),
KZp=ScalarDef(I32),
O=TensorDef(U, S.N, S.OH, S.OW, S.G, S.FG, output=True),
strides=IndexAttrDef(S.SH, S.SW, default=[1, 1]),
dilations=IndexAttrDef(S.DH, S.DW, default=[1, 1]),
):
"""Performs 2-D grouped convolution with zero point offsets.

Layout:
* Input: NHWGC.
* Kernel: GFHWC.

Numeric casting is performed on the operands to the inner multiply, promoting
them to the same data type as the accumulator/output. This includes the zero
point offsets common to quantized operations.
"""
implements(ConvolutionOpInterface)
domain(D.n, D.oh, D.ow, D.g, D.fg, D.kh, D.kw, D.c)
O[D.n, D.oh, D.ow, D.g, D.fg] += (
TypeFn.cast_signed(
U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW, D.g, D.c]
)
- TypeFn.cast_signed(U, IZp)
) * (
TypeFn.cast_signed(U, K[D.g, D.fg, D.kh, D.kw, D.c])
- TypeFn.cast_signed(U, KZp)
)


@linalg_structured_op
def conv_2d_ngchw_gfchw_q(
I=TensorDef(
Expand Down
32 changes: 32 additions & 0 deletions mlir/test/Dialect/Linalg/named-ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -409,6 +409,38 @@ func.func @conv_2d_ngchw_fgchw(%input: memref<?x?x?x?x?xf32>, %filter: memref<?x

// -----

// CHECK-LABEL: func @conv_2d_nhwgc_gfhwc
Copy link
Contributor

@hanhanW hanhanW Oct 25, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you add a test case for static shapes, like below test case? It is not easy to see if the indexing maps are defined properly or not. Having a static shape test case could capture the failure if there is something wrong.

func.func @conv_2d_nhwgc_gfhwc(%input: memref<?x?x?x?x?xf32>, %filter: memref<?x?x?x?x?xf32>, %output: memref<?x?x?x?x?xf32>) {
// CHECK: linalg.conv_2d_nhwgc_gfhwc
// CHECK-SAME: dilations = dense<1> : tensor<2xi64>
// CHECK-SAME: strides = dense<1> : tensor<2xi64>
// CHECK-SAME: ins(%{{.+}}, %{{.+}} : memref<?x?x?x?x?xf32>, memref<?x?x?x?x?xf32>)
// CHECK-SAME: outs(%{{.+}} : memref<?x?x?x?x?xf32>)
linalg.conv_2d_nhwgc_gfhwc {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: memref<?x?x?x?x?xf32>, memref<?x?x?x?x?xf32>)
outs (%output: memref<?x?x?x?x?xf32>)
return
}

// -----

// CHECK-LABEL: func @conv_2d_nhwgc_gfhwc_tensor
func.func @conv_2d_nhwgc_gfhwc_tensor(%input: tensor<1x28x28x2x3xf32>, %filter: tensor<2x8x3x3x3xf32>, %output: tensor<1x26x26x2x8xf32>) -> tensor<1x26x26x2x8xf32> {
// CHECK: linalg.conv_2d_nhwgc_gfhwc
// CHECK-SAME: dilations = dense<1> : tensor<2xi64>
// CHECK-SAME: strides = dense<1> : tensor<2xi64>
// CHECK-SAME: ins(%{{.+}}, %{{.+}} : tensor<1x28x28x2x3xf32>, tensor<2x8x3x3x3xf32>)
// CHECK-SAME: outs(%{{.+}} : tensor<1x26x26x2x8xf32>) -> tensor<1x26x26x2x8xf32>
%0 = linalg.conv_2d_nhwgc_gfhwc {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: tensor<1x28x28x2x3xf32>, tensor<2x8x3x3x3xf32>)
outs (%output: tensor<1x26x26x2x8xf32>) -> tensor<1x26x26x2x8xf32>
return %0 : tensor<1x26x26x2x8xf32>
}

// -----

// CHECK-LABEL: func @conv_2d_ngchw_fgchw_dimensions
func.func @conv_2d_ngchw_fgchw_dimensions(%input: tensor<1x5x3x32x32xf32>, %filter: tensor<2x5x3x3x3xf32>, %init: tensor<1x5x2x30x30xf32>) -> tensor<1x5x2x30x30xf32> {
// CHECK: linalg.conv_2d_ngchw_fgchw
Expand Down
Loading