core: Introduce new atomic datatypes and operation

Add 16-bit and 8-bit floating point datatypes that are often used for machine learning and AI applications. Add atomic diff operation. Signed-off-by: Jianxin Xiong <[email protected]>
ofiwg · Aug 3, 2024 · d3a71f8 · d3a71f8
1 parent 431bec3
commit d3a71f8
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 0 deletions.
diff --git a/include/rdma/fi_domain.h b/include/rdma/fi_domain.h
@@ -226,6 +226,10 @@ enum fi_datatype {
 	FI_LONG_DOUBLE_COMPLEX,
 	FI_INT128,
 	FI_UINT128,
+	FI_FLOAT16,
+	FI_BFLOAT16,
+	FI_FLOAT8_E4M3,
+	FI_FLOAT8_E5M2,
 
 	/* Collective datatypes */
 	FI_VOID = FI_COLLECTIVE_OFFSET,
@@ -251,6 +255,7 @@ enum fi_op {
 	FI_CSWAP_GE,
 	FI_CSWAP_GT,
 	FI_MSWAP,
+	FI_DIFF,
 
 	/* Collective datatypes */
 	FI_NOOP = FI_COLLECTIVE_OFFSET,

diff --git a/man/fi_atomic.3.md b/man/fi_atomic.3.md
@@ -228,6 +228,18 @@ provider implementation constraints.
   a complex number and the second representing the imaginary
   portion.
 
+*FI_FLOAT16*
+: 16-bit half precision floating point value (IEEE 754-2008).
+
+*FI_BFLOAT16*
+: 16-bit brain floating point value (IEEE 754-2008).
+
+*FI_FLOAT8_E4M3*
+: 8-bit floating point value with 4-bit exponent and 3-bit mantissa.
+
+*FI_FLOAT8_E5M2*
+: 8-bit floating point value with 5-bit exponent and 2-bit mantissa.
+
 ## Atomic Operations
 
 The following atomic operations are defined.  An atomic operation
@@ -377,6 +389,13 @@ if (compare[i] > addr[i])
 addr[i] = (buf[i] & compare[i]) | (addr[i] & ~compare[i])
 ```
 
+*FI_DIFF*
+: Calculate the difference
+
+```c
+addr[i] = addr[i] - buf[i]
+```
+
 ## Base Atomic Functions
 
 The base atomic functions -- fi_atomic, fi_atomicv,