[Mlir-commits] [mlir] [mlir][linalg] Add linalg.conv_2d_ngchw_gfchw_q to named ops (PR #92136)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Tue May 14 10:50:25 PDT 2024
https://github.com/zjgarvey updated https://github.com/llvm/llvm-project/pull/92136
>From d5f918b9f0748c19de41dfa0a8a1dbd5f3d07d33 Mon Sep 17 00:00:00 2001
From: zjgarvey <zjgarvey at gmail.com>
Date: Tue, 14 May 2024 17:48:28 +0000
Subject: [PATCH] Add a grouped 2d quantized convolution op
---
.../Linalg/IR/LinalgNamedStructuredOps.yaml | 142 +++++++++++++++++-
.../linalg/opdsl/ops/core_named_ops.py | 34 +++++
2 files changed, 174 insertions(+), 2 deletions(-)
diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml b/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
index 584bfcd8b59dc..5aa8afad0ebf3 100644
--- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
+++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml
@@ -516,7 +516,7 @@ structured_op: !LinalgStructuredOpConfig
--- !LinalgOpConfig
metadata: !LinalgOpMetadata
name: erf
- cpp_class_name: erfOp
+ cpp_class_name: ErfOp
doc: |-
Applies erf(x) elementwise.
@@ -959,7 +959,7 @@ structured_op: !LinalgStructuredOpConfig
--- !LinalgOpConfig
metadata: !LinalgOpMetadata
name: powf
- cpp_class_name: PowFOp
+ cpp_class_name: PowfOp
doc: |-
Takes the powf(lhs, rhs) between two inputs, elementwise. For powf(arg, 2) use `linalg.square`.
@@ -3421,6 +3421,144 @@ structured_op: !LinalgStructuredOpConfig
- !ScalarExpression
scalar_arg: K
--- !LinalgOpConfig
+metadata: !LinalgOpMetadata
+ name: conv_2d_ngchw_gfchw_q
+ cpp_class_name: Conv2DNgchwGfchwQOp
+ doc: |-
+ Performs 2-D grouped convolution with zero-point offsets.
+
+ Layout:
+ * Input: NGCHW.
+ * Kernel: GFCHW.
+
+ 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] ->
+ (s1, s11, s2, s5, s9)>
+ - !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, s11, s3, s7)>
+ - !LinalgOperandDefConfig
+ name: strides
+ 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
+ - !LinalgOperandDefConfig
+ name: dilations
+ kind: index_attr
+ index_attr_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
+ -> (s6, s10)>
+ 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, d5, d3 * s4 + d6 * s6, d4 * s8 + d7 * s10)>
+ - affine_map<(d0, d1, d2, d3, d4, d5, d6, d7)[s0, s1, s2, s3, s4, s5, s6, s7,
+ s8, s9, s10, s11] -> (d1, d2, 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_3d_ndhwc_dhwcf
cpp_class_name: Conv3DNdhwcDhwcfOp
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 ca2bb0c5f7f8a..f1790b1fa2893 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
@@ -937,6 +937,40 @@ def conv_2d_ngchw_gfchw(
U, I[D.n, D.g, D.c, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW]
) * TypeFn.cast_signed(U, K[D.g, D.fg, D.c, D.kh, D.kw])
+ at linalg_structured_op
+def conv_2d_ngchw_gfchw_q(
+ I=TensorDef(
+ T1, S.N, S.G, S.C, S.OH * S.SH + S.KH * S.DH, S.OW * S.SW + S.KW * S.DW
+ ),
+ K=TensorDef(T2, S.G, S.FG, S.C, S.KH, S.KW),
+ IZp=ScalarDef(I32),
+ KZp=ScalarDef(I32),
+ O=TensorDef(U, S.N, S.G, S.FG, S.OH, S.OW, 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: NGCHW.
+ * Kernel: GFCHW.
+
+ 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.g, D.fg, D.oh, D.ow, D.c, D.kh, D.kw)
+ O[D.n, D.g, D.fg, D.oh, D.ow] += (
+ TypeFn.cast_signed(
+ U, I[D.n, D.g, D.c, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW]
+ )
+ - TypeFn.cast_signed(U, IZp)
+ ) * (
+ TypeFn.cast_signed(U, K[D.g, D.fg, D.c, D.kh, D.kw])
+ - TypeFn.cast_signed(U, KZp)
+ )
+
@linalg_structured_op
def conv_3d_ndhwc_dhwcf(
More information about the Mlir-commits
mailing list