Skip to content

Commit 76f2cf9

Browse files
TaoLvTaoLv
committed
doc: graph: op: update supported data types
1 parent df61ffb commit 76f2cf9

File tree

8 files changed

+42
-35
lines changed

8 files changed

+42
-35
lines changed

‎doc/graph/fusion_patterns/sdpa.md

+3-3
Original file line numberDiff line numberDiff line change
@@ -128,9 +128,9 @@ platforms follow the general description in @ref dev_guide_data_types.
128128
4. GPU
129129
- Optimized implementation is available for 4D Q/K/V tensors with shape
130130
defined as (N, H, S, D).
131-
- Optimized implementation is available for floating-point SDPA with `f16`
132-
data type and `D <= 256` on Intel Graphics Products with Intel(R) Xe Matrix
133-
Extensions (Intel(R) XMX) support.
131+
- Optimized implementation is available for `f16` or `bf16` SDPA with `f32`
132+
intermediate data type and `D <= 256` on Intel Graphics Products with
133+
Intel(R) Xe Matrix Extensions (Intel(R) XMX) support.
134134

135135
## Example
136136

‎doc/graph/operations/Add.md

+7-5
Original file line numberDiff line numberDiff line change
@@ -44,8 +44,10 @@ different and auto-broadcasting is allowed if `auto_broadcast` attributes is
4444

4545
Add operation supports the following data type combinations.
4646

47-
| Src_0 / Src_1 | Dst |
48-
|:--------------|:-----|
49-
| f32 | f32 |
50-
| bf16 | bf16 |
51-
| f16 | f16 |
47+
| Src_0 | Src_1 | Dst |
48+
|:----------|:----------|:-----|
49+
| f32 | f32 | f32 |
50+
| bf16 | bf16 | bf16 |
51+
| f16 | f16 | f16 |
52+
| f32 | bf16, f16 | f32 |
53+
| bf16, f16 | f32 | f32 |

‎doc/graph/operations/Divide.md

+7-5
Original file line numberDiff line numberDiff line change
@@ -44,8 +44,10 @@ different and auto-broadcasting is allowed if `auto_broadcast` attributes is
4444

4545
Divide operation supports the following data type combinations.
4646

47-
| Src_0 / Src_1 | Dst |
48-
|:--------------|:-----|
49-
| f32 | f32 |
50-
| bf16 | bf16 |
51-
| f16 | f16 |
47+
| Src_0 | Src_1 | Dst |
48+
|:----------|:----------|:-----|
49+
| f32 | f32 | f32 |
50+
| bf16 | bf16 | bf16 |
51+
| f16 | f16 | f16 |
52+
| f32 | bf16, f16 | f32 |
53+
| bf16, f16 | f32 | f32 |

‎doc/graph/operations/MatMul.md

+5-5
Original file line numberDiff line numberDiff line change
@@ -61,8 +61,8 @@ constructing an operation.
6161

6262
MatMul operation supports the following data type combinations.
6363

64-
| Src | Weights | Bias | Dst |
65-
|:-----|:--------|:-----|:-----|
66-
| f32 | f32 | f32 | f32 |
67-
| bf16 | bf16 | bf16 | bf16 |
68-
| f16 | f16 | f16 | f16 |
64+
| Src | Weights | Bias | Dst |
65+
|:-----|:--------|:-----|:----------|
66+
| f32 | f32 | f32 | f32 |
67+
| bf16 | bf16 | bf16 | f32, bf16 |
68+
| f16 | f16 | f16 | f32, f16 |

‎doc/graph/operations/Multiply.md

+7-5
Original file line numberDiff line numberDiff line change
@@ -44,8 +44,10 @@ different and auto-broadcasting is allowed if `auto_broadcast` attributes is
4444

4545
Multiply operation supports the following data type combinations.
4646

47-
| Src_0 / Src_1 | Dst |
48-
|:--------------|:-----|
49-
| f32 | f32 |
50-
| bf16 | bf16 |
51-
| f16 | f16 |
47+
| Src_0 | Src_1 | Dst |
48+
|:----------|:----------|:-----|
49+
| f32 | f32 | f32 |
50+
| bf16 | bf16 | bf16 |
51+
| f16 | f16 | f16 |
52+
| f32 | bf16, f16 | f32 |
53+
| bf16, f16 | f32 | f32 |

‎doc/graph/operations/Softmax.md

+5-5
Original file line numberDiff line numberDiff line change
@@ -36,8 +36,8 @@ constructing an operation.
3636

3737
SoftMax operation supports the following data type combinations.
3838

39-
| Src | Dst |
40-
|:-----|:-----|
41-
| f32 | f32 |
42-
| bf16 | bf16 |
43-
| f16 | f16 |
39+
| Src | Dst |
40+
|:-----|:----------------|
41+
| f32 | f32, bf16, f16 |
42+
| bf16 | bf16 |
43+
| f16 | f16 |

‎doc/graph/operations/Subtract.md

+7-5
Original file line numberDiff line numberDiff line change
@@ -44,8 +44,10 @@ different and auto-broadcasting is allowed if `auto_broadcast` attributes is
4444

4545
Subtract operation supports the following data type combinations.
4646

47-
| Src_0 / Src_1 | Dst |
48-
|:--------------|:-----|
49-
| f32 | f32 |
50-
| bf16 | bf16 |
51-
| f16 | f16 |
47+
| Src_0 | Src_1 | Dst |
48+
|:----------|:----------|:-----|
49+
| f32 | f32 | f32 |
50+
| bf16 | bf16 | bf16 |
51+
| f16 | f16 | f16 |
52+
| f32 | bf16, f16 | f32 |
53+
| bf16, f16 | f32 | f32 |

‎doc/graph/programming_model/low_precision.md

+1-2
Original file line numberDiff line numberDiff line change
@@ -52,7 +52,6 @@ Graph operations support bf16 and f16 data types.
5252

5353
A TypeCast operation performing down conversion should be inserted clearly to
5454
indicate the use of low numeric precision. oneDNN Graph implementation fully
55-
honors the API-specified numeric precision and only performs the computation
56-
using the API-specified or higher numeric precision.
55+
honors the API-specified numeric precision.
5756

5857
@img{bf16_programming.jpg,Figure 2: Overview of bf16 programming model.,80%,}

0 commit comments

Comments
 (0)