[Mlir-commits] [mlir] Quantile Type and Low FP Support (PR #190321)

[llvmlistbot at llvm.org]

================
@@ -1109,6 +1109,94 @@ def Builtin_Opaque : Builtin_Type<"Opaque", "opaque"> {
   let genVerifyDecl = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// QuantileType
+//===----------------------------------------------------------------------===//
+
+def Builtin_Quantile : Builtin_Type<"Quantile", "quantile",
+    [QuantStorageTypeInterface, MemRefElementTypeInterface]> {
+  let summary = "Quantile-based type with a lookup table of quantile values";
+  let description = [{
+    Syntax:
+
+    ```
+    quantile-type ::= `quantile` `<` type ( `<` int `,` int `>` )? `:` type `,` `{` float-list `}` `>`
+    ```
+
+    A quantile type represents a quantile-based floating point encoding, where
+    discrete storage values map to floating-point values via a quantile lookup
+    table. The type has a storage type (how raw values are stored, e.g., `ui4`,
+    `si8`), a quantile type (the expressed floating-point precision, e.g.,
+    `f16`, `f32`), and an array of quantile values that define the mapping.
+
+    Optionally, explicit minimum and maximum storage values can be specified
+    after the storage type as `<min,max>`, overriding the defaults derived from
+    the storage type width.
+
+    This type is used for weight compression schemes like NF4 (NormalFloat4)
----------------
Roman-Pevnyi wrote:

[Normalized float4](https://github.com/openvinotoolkit/npu_compiler/blob/90b6098b9ee96055d633dc520354434bae22e336/src/vpux_compiler/include/vpux/compiler/core/types/quantile_float/types.hpp#L62)

https://github.com/llvm/llvm-project/pull/190321