[Mlir-commits] [mlir] [mlir][linalg] Add Grouped Convolution Ops: conv_2d_nhwgc_gfhwc and conv_2d_nhwgc_gfhwc_q (PR #108192)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Wed Oct 30 03:01:10 PDT 2024


https://github.com/stefankoncarevic updated https://github.com/llvm/llvm-project/pull/108192

>From 4aed02344f3ae3313008b54023f341a64a1114f6 Mon Sep 17 00:00:00 2001
From: Stefan Koncarevic <skoncare at amd.com>
Date: Thu, 19 Sep 2024 12:51:40 +0000
Subject: [PATCH] This commit introduces two new Linalg operations:
 `conv_2d_nhwgc_gfhwc` and `conv_2d_nhwgc_gfhwc_q`. These operations perform
 2-D grouped convolutions with and without zero point offsets, respectively.
 The input layout is NHWGC, and the kernel layout is GFHWC. These additions
 enhance support for grouped convolution operations in MLIR.

---
 .../Linalg/IR/LinalgNamedStructuredOps.yaml   | 237 ++++++++++++++++++
 .../linalg/opdsl/ops/core_named_ops.py        |  61 +++++
 mlir/test/Dialect/Linalg/named-ops.mlir       |  32 +++
 3 files changed, 330 insertions(+)

diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml b/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
index bf2f26de26e9ed..8cd63bc9270759 100644
--- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
+++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
@@ -3547,6 +3547,243 @@ 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
diff --git a/mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py b/mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py
index b45fecd0ee1457..89895760cad743 100644
--- a/mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py
+++ b/mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py
@@ -981,6 +981,67 @@ def conv_2d_ngchw_gfchw(
     ) * TypeFn.cast_signed(U, K[D.g, D.fg, D.c, D.kh, D.kw])
 
 
+ at 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])
+
+
+ at 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(
diff --git a/mlir/test/Dialect/Linalg/named-ops.mlir b/mlir/test/Dialect/Linalg/named-ops.mlir
index 02ecbed232c8b5..bc0cfd52e8b51a 100644
--- a/mlir/test/Dialect/Linalg/named-ops.mlir
+++ b/mlir/test/Dialect/Linalg/named-ops.mlir
@@ -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
+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



More information about the Mlir-commits mailing list