Merge pull request #9996 from vespa-engine/havardpe/propagate-float-cell-type-and-avoid-using-it-too-much-right-now

propagate float and stop using it too much

10 files changed, 100 insertions, 30 deletions

diff --git a/config-model/src/main/javacc/SDParser.jj b/config-model/src/main/javacc/SDParser.jj
index 6dde12f0fac..3141f7f7164 100644
--- a/config-model/src/main/javacc/SDParser.jj
+++ b/config-model/src/main/javacc/SDParser.jj
@@ -297,7 +297,7 @@ TOKEN :
 | < MUTABLE: "mutable" >
 | < FASTSEARCH: "fast-search" >
 | < HUGE: "huge" >
-| < TENSOR_TYPE: "tensor(" (~["(",")"])+ ")" >
+| < TENSOR_TYPE: "tensor" ("<" (~["<",">"])+ ">")? "(" (~["(",")"])+ ")" >
 | < TENSOR_VALUE_SL: "value" (" ")* ":" (" ")* ("{"<BRACE_SL_LEVEL_1>) ("\n")? >
 | < TENSOR_VALUE_ML: "value" (<SEARCHLIB_SKIP>)? "{" (["\n"," "])* ("{"<BRACE_ML_LEVEL_1>) (["\n"," "])* "}" ("\n")? >
 | < COMPRESSION: "compression" >
diff --git a/config-model/src/test/derived/tensor/attributes.cfg b/config-model/src/test/derived/tensor/attributes.cfg
index 4b54e67f8b8..f1c95da7084 100644
--- a/config-model/src/test/derived/tensor/attributes.cfg
+++ b/config-model/src/test/derived/tensor/attributes.cfg
@@ -61,3 +61,24 @@ attribute[].upperbound 9223372036854775807
 attribute[].densepostinglistthreshold 0.4
 attribute[].tensortype "tensor(x[10],y[20])"
 attribute[].imported false
+attribute[].name "f5"
+attribute[].datatype TENSOR
+attribute[].collectiontype SINGLE
+attribute[].removeifzero false
+attribute[].createifnonexistent false
+attribute[].fastsearch false
+attribute[].huge false
+attribute[].ismutable false
+attribute[].sortascending true
+attribute[].sortfunction UCA
+attribute[].sortstrength PRIMARY
+attribute[].sortlocale ""
+attribute[].enablebitvectors false
+attribute[].enableonlybitvector false
+attribute[].fastaccess false
+attribute[].arity 8
+attribute[].lowerbound -9223372036854775808
+attribute[].upperbound 9223372036854775807
+attribute[].densepostinglistthreshold 0.4
+attribute[].tensortype "tensor<float>(x[10])"
+attribute[].imported false
diff --git a/config-model/src/test/derived/tensor/documenttypes.cfg b/config-model/src/test/derived/tensor/documenttypes.cfg
index 3b85e9b3a58..0b75644f4ae 100644
--- a/config-model/src/test/derived/tensor/documenttypes.cfg
+++ b/config-model/src/test/derived/tensor/documenttypes.cfg
@@ -40,6 +40,11 @@ documenttype[].datatype[].sstruct.field[].id 1224191509
 documenttype[].datatype[].sstruct.field[].id_v6 1039544782
 documenttype[].datatype[].sstruct.field[].datatype 21
 documenttype[].datatype[].sstruct.field[].detailedtype "tensor(x[10],y[20])"
+documenttype[].datatype[].sstruct.field[].name "f5"
+documenttype[].datatype[].sstruct.field[].id 329055840
+documenttype[].datatype[].sstruct.field[].id_v6 105711688
+documenttype[].datatype[].sstruct.field[].datatype 21
+documenttype[].datatype[].sstruct.field[].detailedtype "tensor<float>(x[10])"
 documenttype[].datatype[].id -1903234535
 documenttype[].datatype[].type STRUCT
 documenttype[].datatype[].array.element.id 0
@@ -59,3 +64,4 @@ documenttype[].fieldsets{[document]}.fields[] "f1"
 documenttype[].fieldsets{[document]}.fields[] "f2"
 documenttype[].fieldsets{[document]}.fields[] "f3"
 documenttype[].fieldsets{[document]}.fields[] "f4"
+documenttype[].fieldsets{[document]}.fields[] "f5"
diff --git a/config-model/src/test/derived/tensor/rank-profiles.cfg b/config-model/src/test/derived/tensor/rank-profiles.cfg
index 471343da63c..87c0b6fab42 100644
--- a/config-model/src/test/derived/tensor/rank-profiles.cfg
+++ b/config-model/src/test/derived/tensor/rank-profiles.cfg
@@ -5,6 +5,8 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f3"
 rankprofile[].fef.property[].value "tensor(x{})"
 rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
 rankprofile[].name "unranked"
 rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "value(0)"
@@ -20,6 +22,8 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f3"
 rankprofile[].fef.property[].value "tensor(x{})"
 rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
 rankprofile[].name "profile1"
 rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
@@ -31,6 +35,8 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f3"
 rankprofile[].fef.property[].value "tensor(x{})"
 rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
 rankprofile[].name "profile2"
 rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
@@ -42,6 +48,8 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f3"
 rankprofile[].fef.property[].value "tensor(x{})"
 rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
 rankprofile[].name "profile3"
 rankprofile[].fef.property[].name "rankingExpression(joinedtensors).rankingScript"
 rankprofile[].fef.property[].value "tensor(i[10])(i) * attribute(f4)"
@@ -57,3 +65,18 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f3"
 rankprofile[].fef.property[].value "tensor(x{})"
 rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
+rankprofile[].name "profile4"
+rankprofile[].fef.property[].name "vespa.rank.firstphase"
+rankprofile[].fef.property[].value "rankingExpression(firstphase)"
+rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
+rankprofile[].fef.property[].value "reduce(attribute(f5), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f2"
+rankprofile[].fef.property[].value "tensor(x[2],y[])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f3"
+rankprofile[].fef.property[].value "tensor(x{})"
+rankprofile[].fef.property[].name "vespa.type.attribute.f4"
+rankprofile[].fef.property[].value "tensor(x[10],y[20])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
diff --git a/config-model/src/test/derived/tensor/summary.cfg b/config-model/src/test/derived/tensor/summary.cfg
index 5e5507836b0..903b6033297 100644
--- a/config-model/src/test/derived/tensor/summary.cfg
+++ b/config-model/src/test/derived/tensor/summary.cfg
@@ -1,5 +1,5 @@
-defaultsummaryid 289405525
-classes[].id 289405525
+defaultsummaryid 898020074
+classes[].id 898020074
 classes[].name "default"
 classes[].fields[].name "f1"
 classes[].fields[].type "tensor"
@@ -7,13 +7,15 @@ classes[].fields[].name "f3"
 classes[].fields[].type "tensor"
 classes[].fields[].name "f4"
 classes[].fields[].type "tensor"
+classes[].fields[].name "f5"
+classes[].fields[].type "tensor"
 classes[].fields[].name "rankfeatures"
 classes[].fields[].type "featuredata"
 classes[].fields[].name "summaryfeatures"
 classes[].fields[].type "featuredata"
 classes[].fields[].name "documentid"
 classes[].fields[].type "longstring"
-classes[].id 1860420340
+classes[].id 193983608
 classes[].name "attributeprefetch"
 classes[].fields[].name "f2"
 classes[].fields[].type "tensor"
@@ -21,6 +23,8 @@ classes[].fields[].name "f3"
 classes[].fields[].type "tensor"
 classes[].fields[].name "f4"
 classes[].fields[].type "tensor"
+classes[].fields[].name "f5"
+classes[].fields[].type "tensor"
 classes[].fields[].name "rankfeatures"
 classes[].fields[].type "featuredata"
 classes[].fields[].name "summaryfeatures"
diff --git a/config-model/src/test/derived/tensor/tensor.sd b/config-model/src/test/derived/tensor/tensor.sd
index e3e8cf43347..622be033229 100644
--- a/config-model/src/test/derived/tensor/tensor.sd
+++ b/config-model/src/test/derived/tensor/tensor.sd
@@ -8,12 +8,15 @@ search tensor {
     field f2 type tensor(x[2],y[]) {
       indexing: attribute
     }
-    field f3 type tensor(x{}) {
+    field f3 type tensor<double>(x{}) {
       indexing: attribute | summary
     }
     field f4 type tensor(x[10],y[20]) {
       indexing: attribute | summary
     }
+    field f5 type tensor<float>(x[10]) {
+      indexing: attribute | summary
+    }
   }
   
   rank-profile profile1 {
@@ -44,4 +47,12 @@ search tensor {
 
   }
 
+  rank-profile profile4 {
+
+    first-phase {
+      expression: sum(attribute(f5))
+    }
+  
+  }
+
 }
diff --git a/model-integration/src/main/java/ai/vespa/rankingexpression/importer/onnx/TypeConverter.java b/model-integration/src/main/java/ai/vespa/rankingexpression/importer/onnx/TypeConverter.java
index 29d600fa7c6..8c9fe60e1d4 100644
--- a/model-integration/src/main/java/ai/vespa/rankingexpression/importer/onnx/TypeConverter.java
+++ b/model-integration/src/main/java/ai/vespa/rankingexpression/importer/onnx/TypeConverter.java
@@ -53,17 +53,17 @@ class TypeConverter {
 
     private static TensorType.Value toValueType(Onnx.TensorProto.DataType dataType) {
         switch (dataType) {
-            case FLOAT: return TensorType.Value.FLOAT;
+            case FLOAT: return TensorType.Value.DOUBLE;
             case DOUBLE: return TensorType.Value.DOUBLE;
             // Imperfect conversion, for now:
-            case BOOL: return TensorType.Value.FLOAT;
-            case INT8: return TensorType.Value.FLOAT;
-            case INT16: return TensorType.Value.FLOAT;
-            case INT32: return TensorType.Value.FLOAT;
+            case BOOL: return TensorType.Value.DOUBLE;
+            case INT8: return TensorType.Value.DOUBLE;
+            case INT16: return TensorType.Value.DOUBLE;
+            case INT32: return TensorType.Value.DOUBLE;
             case INT64: return TensorType.Value.DOUBLE;
-            case UINT8: return TensorType.Value.FLOAT;
-            case UINT16: return TensorType.Value.FLOAT;
-            case UINT32: return TensorType.Value.FLOAT;
+            case UINT8: return TensorType.Value.DOUBLE;
+            case UINT16: return TensorType.Value.DOUBLE;
+            case UINT32: return TensorType.Value.DOUBLE;
             case UINT64: return TensorType.Value.DOUBLE;
             default: throw new IllegalArgumentException("A ONNX tensor with data type " + dataType +
                                                         " cannot be converted to a Vespa tensor type");
diff --git a/model-integration/src/main/java/ai/vespa/rankingexpression/importer/tensorflow/TypeConverter.java b/model-integration/src/main/java/ai/vespa/rankingexpression/importer/tensorflow/TypeConverter.java
index d8ddb01b650..08c0564ed8a 100644
--- a/model-integration/src/main/java/ai/vespa/rankingexpression/importer/tensorflow/TypeConverter.java
+++ b/model-integration/src/main/java/ai/vespa/rankingexpression/importer/tensorflow/TypeConverter.java
@@ -85,19 +85,19 @@ class TypeConverter {
 
     private static TensorType.Value toValueType(DataType dataType) {
         switch (dataType) {
-            case DT_FLOAT: return TensorType.Value.FLOAT;
+            case DT_FLOAT: return TensorType.Value.DOUBLE;
             case DT_DOUBLE: return TensorType.Value.DOUBLE;
             // Imperfect conversion, for now:
-            case DT_BOOL: return TensorType.Value.FLOAT;
-            case DT_BFLOAT16: return TensorType.Value.FLOAT;
-            case DT_HALF: return TensorType.Value.FLOAT;
-            case DT_INT8: return TensorType.Value.FLOAT;
-            case DT_INT16: return TensorType.Value.FLOAT;
-            case DT_INT32: return TensorType.Value.FLOAT;
+            case DT_BOOL: return TensorType.Value.DOUBLE;
+            case DT_BFLOAT16: return TensorType.Value.DOUBLE;
+            case DT_HALF: return TensorType.Value.DOUBLE;
+            case DT_INT8: return TensorType.Value.DOUBLE;
+            case DT_INT16: return TensorType.Value.DOUBLE;
+            case DT_INT32: return TensorType.Value.DOUBLE;
             case DT_INT64: return TensorType.Value.DOUBLE;
-            case DT_UINT8: return TensorType.Value.FLOAT;
-            case DT_UINT16: return TensorType.Value.FLOAT;
-            case DT_UINT32: return TensorType.Value.FLOAT;
+            case DT_UINT8: return TensorType.Value.DOUBLE;
+            case DT_UINT16: return TensorType.Value.DOUBLE;
+            case DT_UINT32: return TensorType.Value.DOUBLE;
             case DT_UINT64: return TensorType.Value.DOUBLE;
             default: throw new IllegalArgumentException("A TensorFlow tensor with data type " + dataType +
                                                         " cannot be converted to a Vespa tensor type");
@@ -106,12 +106,12 @@ class TypeConverter {
 
     private static TensorType.Value toValueType(org.tensorflow.DataType dataType) {
         switch (dataType) {
-            case FLOAT: return TensorType.Value.FLOAT;
+            case FLOAT: return TensorType.Value.DOUBLE;
             case DOUBLE: return TensorType.Value.DOUBLE;
             // Imperfect conversion, for now:
-            case BOOL: return TensorType.Value.FLOAT;
-            case INT32: return TensorType.Value.FLOAT;
-            case UINT8: return TensorType.Value.FLOAT;
+            case BOOL: return TensorType.Value.DOUBLE;
+            case INT32: return TensorType.Value.DOUBLE;
+            case UINT8: return TensorType.Value.DOUBLE;
             case INT64: return TensorType.Value.DOUBLE;
             default: throw new IllegalArgumentException("A TensorFlow tensor with data type " + dataType +
                                                         " cannot be converted to a Vespa tensor type");
diff --git a/model-integration/src/test/java/ai/vespa/rankingexpression/importer/onnx/OnnxMnistSoftmaxImportTestCase.java b/model-integration/src/test/java/ai/vespa/rankingexpression/importer/onnx/OnnxMnistSoftmaxImportTestCase.java
index 07814687dc6..424e4d6c57c 100644
--- a/model-integration/src/test/java/ai/vespa/rankingexpression/importer/onnx/OnnxMnistSoftmaxImportTestCase.java
+++ b/model-integration/src/test/java/ai/vespa/rankingexpression/importer/onnx/OnnxMnistSoftmaxImportTestCase.java
@@ -43,14 +43,14 @@ public class OnnxMnistSoftmaxImportTestCase {
         // Check inputs
         assertEquals(1, model.inputs().size());
         assertTrue(model.inputs().containsKey("Placeholder"));
-        assertEquals(TensorType.fromSpec("tensor<float>(d0[],d1[784])"), model.inputs().get("Placeholder"));
+        assertEquals(TensorType.fromSpec("tensor(d0[],d1[784])"), model.inputs().get("Placeholder"));
 
         // Check signature
         ImportedMlFunction output = model.defaultSignature().outputFunction("add", "add");
         assertNotNull(output);
         assertEquals("join(reduce(join(rename(Placeholder, (d0, d1), (d0, d2)), constant(test_Variable), f(a,b)(a * b)), sum, d2), constant(test_Variable_1), f(a,b)(a + b))",
                      output.expression());
-        assertEquals(TensorType.fromSpec("tensor<float>(d0[],d1[784])"),
+        assertEquals(TensorType.fromSpec("tensor(d0[],d1[784])"),
                      model.inputs().get(model.defaultSignature().inputs().get("Placeholder")));
         assertEquals("{Placeholder=tensor(d0[],d1[784])}", output.argumentTypes().toString());
     }
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/TensorType.java b/vespajlib/src/main/java/com/yahoo/tensor/TensorType.java
index 8e566fac0b6..9869f1e908c 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/TensorType.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/TensorType.java
@@ -168,7 +168,11 @@ public class TensorType {
 
     @Override
     public String toString() {
-        return "tensor(" + dimensions.stream().map(Dimension::toString).collect(Collectors.joining(",")) + ")";
+        if ((rank() == 0) || (valueType == Value.DOUBLE)) { 
+            return "tensor(" + dimensions.stream().map(Dimension::toString).collect(Collectors.joining(",")) + ")";
+        } else {
+            return "tensor<" + valueType + ">(" + dimensions.stream().map(Dimension::toString).collect(Collectors.joining(",")) + ")";
+        }
     }
 
     @Override
@@ -177,6 +181,7 @@ public class TensorType {
         if (o == null || getClass() != o.getClass()) return false;
 
         TensorType other = (TensorType)o;
+        if ( (this.rank() == 0) && (other.rank() == 0)) return true;
         if ( this.valueType != other.valueType) return false;
         if ( ! this.dimensions.equals(other.dimensions)) return false;
         return true;