Support F32_16 precision

litert/c/litert_common.h

@@ -309,6 +309,9 @@ typedef enum {
   kLiteRtDelegatePrecisionDefault = 0,
   kLiteRtDelegatePrecisionFp16 = 1,
   kLiteRtDelegatePrecisionFp32 = 2,
+  // Use FP16 except for some cases where FP32 is needed, e.g. accumulation
+  // in convolution.
+  kLiteRtDelegatePrecisionFp32_16 = 3,
 } LiteRtDelegatePrecision;
 
 typedef enum {

litert/cc/options/litert_gpu_options.h

@@ -68,6 +68,7 @@ class GpuOptions {
     kDefault = kLiteRtDelegatePrecisionDefault,
     kFp16 = kLiteRtDelegatePrecisionFp16,
     kFp32 = kLiteRtDelegatePrecisionFp32,
+    kFp32_16 = kLiteRtDelegatePrecisionFp32_16,
   };
   Expected<void> SetPrecision(Precision precision);
 

litert/kotlin/src/main/jni/litert_compiled_model_jni.cc

@@ -122,6 +122,7 @@ enum Precision {
   kPrecisionDefault = 0,
   kPrecisionFp16 = 1,
   kPrecisionFp32 = 2,
+  kPrecisionFp32_16 = 3,
 };
 
 // Converts the precision string to LiteRtDelegatePrecision.
@@ -132,6 +133,8 @@ GpuOptions::Precision ToGpuOptionsPrecision(const char* precision_str) {
       return GpuOptions::Precision::kFp16;
     case kPrecisionFp32:
       return GpuOptions::Precision::kFp32;
+    case kPrecisionFp32_16:
+      return GpuOptions::Precision::kFp32_16;
     default:
       return GpuOptions::Precision::kDefault;
   }

litert/kotlin/src/main/kotlin/com/google/ai/edge/litert/Model.kt

@@ -114,6 +114,7 @@ private constructor(
       DEFAULT(0),
       FP16(1),
       FP32(2),
+      FP32_16(3),
     }
 
     /** Buffer storage type for GPU options. */