Replicate the scalar scaling factor for 4-bit tensor-wise quantized weights to the number of channels, as is done for 8-bit weights.

include/xnnpack.h

@@ -309,6 +309,9 @@ enum xnn_datatype {
   /// Quantized 2-bit signed integer with shared per-channel quantization
   /// parameters, but packed into 8-bit integers.
   xnn_datatype_qcint2 = 18,
+  /// Quantized 4-bit signed integer with shared per-Value quantization
+  /// parameters.
+  xnn_datatype_qint4 = 19,
 };
 
 /// Define a tensor-type Value and add it to a Subgraph.

src/datatype.c

@@ -22,6 +22,7 @@ bool xnn_datatype_is_real(enum xnn_datatype t) {
     case xnn_datatype_qcint8:
     case xnn_datatype_qcint32:
     case xnn_datatype_qcint4:
+    case xnn_datatype_qint4:
     case xnn_datatype_qcint2:
     case xnn_datatype_qdint8:
     case xnn_datatype_qduint8:
@@ -48,6 +49,7 @@ bool xnn_datatype_is_integral(enum xnn_datatype t) {
     case xnn_datatype_qcint8:
     case xnn_datatype_qcint32:
     case xnn_datatype_qcint4:
+    case xnn_datatype_qint4:
     case xnn_datatype_qcint2:
     case xnn_datatype_qdint8:
     case xnn_datatype_qduint8:
@@ -72,6 +74,7 @@ bool xnn_datatype_is_quantized(enum xnn_datatype t) {
     case xnn_datatype_qcint8:
     case xnn_datatype_qcint32:
     case xnn_datatype_qcint4:
+    case xnn_datatype_qint4:
     case xnn_datatype_qcint2:
     case xnn_datatype_qdint8:
     case xnn_datatype_qduint8:
@@ -119,6 +122,7 @@ size_t xnn_datatype_log2_size_bits(enum xnn_datatype t) {
       return -1;
     case xnn_datatype_qcint2:
       return 1;
+    case xnn_datatype_qint4:
     case xnn_datatype_qcint4:
     case xnn_datatype_qbint4:
       return 2;
@@ -172,6 +176,7 @@ bool xnn_datatype_is_byte_addressable(enum xnn_datatype t) {
     case xnn_datatype_fp16:
     case xnn_datatype_bf16:
     case xnn_datatype_qint8:
+    case xnn_datatype_qint4:
     case xnn_datatype_pqint8:
     case xnn_datatype_quint8:
     case xnn_datatype_qint32:

src/enums/datatype-strings.c

@@ -52,6 +52,8 @@ const char* xnn_datatype_to_string(enum xnn_datatype type) {
       return "QBINT4";
     case xnn_datatype_qcint2:
       return "QCINT2";
+    case xnn_datatype_qint4:
+      return "QINT4";
   }
   XNN_UNREACHABLE;
   return NULL;

src/subgraph.c

@@ -3727,7 +3727,7 @@ static enum xnn_status optimize_common_subgraphs_iter(
         // is static.
         if (xnn_shape_multiply_all_dims(
                 &node->params.static_reshape.new_shape) != 0) {
-          xnn_log_info(
+          xnn_log_debug(
               "Marking output of static_reshape[#%u](v%03u) as static shaped.",
               node->id, node->inputs[0]);
           subgraph->values[node->outputs[0]].shape =
@@ -3796,7 +3796,7 @@ enum xnn_status xnn_subgraph_optimize_common_subgraphs(
       if (xnn_shape_multiply_all_dims(&value->shape) == 1) {
         // Get the value as a float.
         const float value_as_float = get_scalar_value_as_float(value);
-        xnn_log_info("v%03u is a constant: %e.", value->id, value_as_float);
+        xnn_log_debug("v%03u is a constant: %e.", value->id, value_as_float);
 
         // Mark the value accordingly.
         value->flags |= (value_as_float == 0.0f)   ? XNN_VALUE_FLAG_IS_ZERO

src/tensor.c

@@ -61,6 +61,7 @@ static enum xnn_status check_zero_point(
         return xnn_status_invalid_parameter;
       }
       break;
+    case xnn_datatype_qint4:
     case xnn_datatype_qint8:
       if ((int32_t) (int8_t) zero_point != zero_point) {
         xnn_log_error(

ynnpack/xnnpack/utils.cc

@@ -1202,6 +1202,7 @@ ynn_type type_from_xnn(xnn_datatype type) {
       return ynn_type_bf16;
     case xnn_datatype_qduint8:
       return ynn_type_uint8;
+    case xnn_datatype_qint4:
     case xnn_datatype_qpint8:
     case xnn_datatype_pfp32:
     case xnn_datatype_pfp16: