Ensure int64 accumulation for bias-less 16x8 ops in TFLM kernels (#3598)

veblush · web-flow · commit 7ca66d18e748 · 2026-06-16T23:58:19.000Z
diff --git a/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc b/tensorflow/lite/micro/kernels/cmsis_nn/conv.cc
@@ -392,7 +392,7 @@ TfLiteStatus EvalInt16x8(TfLiteContext* context, TfLiteNode* node) {
   }
   const bool requires_int32_accum =
       (bias != nullptr && bias->type == kTfLiteInt32) ||
-      (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);
+      (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);
   if (requires_int32_accum) {
     return EvalQuantizedPerChannel<int16_t, int32_t, kTfLiteInt16>(
         context, node, params, data, input, filter, bias, output);
diff --git a/tensorflow/lite/micro/kernels/conv.cc b/tensorflow/lite/micro/kernels/conv.cc
@@ -88,7 +88,7 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {
       }
       const bool requires_int32_accum =
           (bias != nullptr && bias->type == kTfLiteInt32) ||
-          (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);
+          (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);
       if (requires_int32_accum) {
         reference_integer_ops::ConvPerChannel(
             ConvParamsQuantized(params, data),
diff --git a/tensorflow/lite/micro/kernels/fully_connected.cc b/tensorflow/lite/micro/kernels/fully_connected.cc
@@ -240,7 +240,7 @@ TfLiteStatus FullyConnectedEval(TfLiteContext* context, TfLiteNode* node) {
         case kTfLiteInt8: {
           const bool requires_int32_accum =
               (bias != nullptr && bias->type == kTfLiteInt32) ||
-              (bias == nullptr && params->quantized_bias_type != kTfLiteInt64);
+              (bias == nullptr && params->quantized_bias_type == kTfLiteInt32);
           if (requires_int32_accum) {
             data.is_per_channel
                 ? tflite::reference_integer_ops::FullyConnectedPerChannel(
diff --git a/tensorflow/lite/micro/kernels/xtensa/conv.cc b/tensorflow/lite/micro/kernels/xtensa/conv.cc
@@ -117,7 +117,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
       // implementation, production use-cases should only have int64 bias.
       const bool requires_int32_accum =
           (bias != nullptr && bias->type == kTfLiteInt32) ||
-          (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);
+          (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);
       if (requires_int32_accum) {
         return ConvReferenceEvalInt16(context, node);
       } else {
diff --git a/tensorflow/lite/micro/kernels/xtensa/conv_int16_reference.cc b/tensorflow/lite/micro/kernels/xtensa/conv_int16_reference.cc
@@ -64,7 +64,7 @@ TfLiteStatus ConvReferenceEvalInt16(TfLiteContext* context, TfLiteNode* node) {
   }
   const bool requires_int32_accum =
       (bias != nullptr && bias->type == kTfLiteInt32) ||
-      (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);
+      (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);
 
   if (requires_int32_accum) {
     reference_integer_ops::ConvPerChannel(

Original file line number	Diff line number	Diff line change
`@@ -392,7 +392,7 @@ TfLiteStatus EvalInt16x8(TfLiteContext* context, TfLiteNode* node) {`
`392`	`392`	`}`
`393`	`393`	`const bool requires_int32_accum =`
`394`	`394`	`(bias != nullptr && bias->type == kTfLiteInt32) \|\|`
`395`		`- (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);`
	`395`	`+ (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);`
`396`	`396`	`if (requires_int32_accum) {`
`397`	`397`	`return EvalQuantizedPerChannel<int16_t, int32_t, kTfLiteInt16>(`
`398`	`398`	`context, node, params, data, input, filter, bias, output);`
Original file line number	Diff line number	Diff line change
`@@ -88,7 +88,7 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {`
`88`	`88`	`}`
`89`	`89`	`const bool requires_int32_accum =`
`90`	`90`	`(bias != nullptr && bias->type == kTfLiteInt32) \|\|`
`91`		`- (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);`
	`91`	`+ (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);`
`92`	`92`	`if (requires_int32_accum) {`
`93`	`93`	`reference_integer_ops::ConvPerChannel(`
`94`	`94`	`ConvParamsQuantized(params, data),`
Original file line number	Diff line number	Diff line change
`@@ -64,7 +64,7 @@ TfLiteStatus ConvReferenceEvalInt16(TfLiteContext* context, TfLiteNode* node) {`
`64`	`64`	`}`
`65`	`65`	`const bool requires_int32_accum =`
`66`	`66`	`(bias != nullptr && bias->type == kTfLiteInt32) \|\|`
`67`		`- (bias == nullptr && params.quantized_bias_type != kTfLiteInt64);`
	`67`	`+ (bias == nullptr && params.quantized_bias_type == kTfLiteInt32);`
`68`	`68`
`69`	`69`	`if (requires_int32_accum) {`
`70`	`70`	`reference_integer_ops::ConvPerChannel(`