From a9f0624d6b2f735a13d7d21aba913048c49d5c25 Mon Sep 17 00:00:00 2001 From: Philipp Zabel Date: Tue, 26 Nov 2024 16:47:47 +0100 Subject: [PATCH] teflon: Reject per-axis quantization Until a workaround for missing hardware support is implemented, stop pretending to support convolution operations on tensors with per-axis quantization. Reviewed-by: Tomeu Vizoso Signed-off-by: Philipp Zabel Part-of: --- src/gallium/frontends/teflon/tfl_device.c | 39 ++++++++++++++++++++--- 1 file changed, 35 insertions(+), 4 deletions(-) diff --git a/src/gallium/frontends/teflon/tfl_device.c b/src/gallium/frontends/teflon/tfl_device.c index 7ecf15ba694..583a71b6e91 100644 --- a/src/gallium/frontends/teflon/tfl_device.c +++ b/src/gallium/frontends/teflon/tfl_device.c @@ -390,6 +390,21 @@ partition_invoke(TfLiteContext *tf_context, TfLiteNode *node) return kTfLiteOk; } +static bool +tensor_quantization_supported(TfLiteTensor *tensor) +{ + if (tensor->quantization.type == kTfLiteAffineQuantization) { + TfLiteAffineQuantization *affine = (TfLiteAffineQuantization *)tensor->quantization.params; + + /* + * Per-axis quantization not supported, for details see: + * https://ai.google.dev/edge/litert/models/quantization_spec#per-axis_vs_per-tensor + */ + return affine->scale->size == 1 && affine->zero_point->size == 1; + } + return false; +} + static TfLiteStatus PrepareDelegate(TfLiteContext *context, TfLiteDelegate *delegate) { @@ -410,10 +425,18 @@ PrepareDelegate(TfLiteContext *context, TfLiteDelegate *delegate) switch(registration->builtin_code) { case kTfLiteBuiltinConv2d: { + TfLiteTensor *input_tensor = &context->tensors[node->inputs->data[0]]; + TfLiteTensor *weight_tensor = &context->tensors[node->inputs->data[1]]; + TfLiteTensor *bias_tensor = &context->tensors[node->inputs->data[2]]; + TfLiteTensor *output_tensor = &context->tensors[node->outputs->data[0]]; TfLiteConvParams* params = (TfLiteConvParams*)node->builtin_data; - // Dilation not yet implemented - if ((params->activation == kTfLiteActNone || + // Dilation and per-axis quantization not yet implemented + if (tensor_quantization_supported(input_tensor) && + tensor_quantization_supported(weight_tensor) && + tensor_quantization_supported(bias_tensor) && + tensor_quantization_supported(output_tensor) && + (params->activation == kTfLiteActNone || params->activation == kTfLiteActRelu) && (registration->version < 2 || (params->dilation_width_factor == 1 && @@ -423,10 +446,18 @@ PrepareDelegate(TfLiteContext *context, TfLiteDelegate *delegate) break; } case kTfLiteBuiltinDepthwiseConv2d: { + TfLiteTensor *input_tensor = &context->tensors[node->inputs->data[0]]; + TfLiteTensor *weight_tensor = &context->tensors[node->inputs->data[1]]; + TfLiteTensor *bias_tensor = &context->tensors[node->inputs->data[2]]; + TfLiteTensor *output_tensor = &context->tensors[node->outputs->data[0]]; TfLiteDepthwiseConvParams* params = (TfLiteDepthwiseConvParams*)node->builtin_data; - // Dilation not yet implemented - if ((params->activation == kTfLiteActNone || + // Dilation and per-axis quantization not yet implemented + if (tensor_quantization_supported(input_tensor) && + tensor_quantization_supported(weight_tensor) && + tensor_quantization_supported(bias_tensor) && + tensor_quantization_supported(output_tensor) && + (params->activation == kTfLiteActNone || params->activation == kTfLiteActRelu) && (registration->version < 2 || (params->dilation_width_factor == 1 &&