-
Notifications
You must be signed in to change notification settings - Fork 684
Description
🚀 The feature, motivation and pitch
Infer / Run mv2 and mv3 on FVP with fused ops + cmsis-nn integration.
To fully support it, we need to support the following essential ops
- Op Support
Quantized Arithmetic Operations
cortex_m::quantized_add.out — Element-wise addition using ARM CMSIS-NN function arm_elementwise_add_s8()
cortex_m::quantized_mul.out — Element-wise multiplication
Neural Network Layers / MatMul
cortex_m::quantized_conv2d.out — Convolution using arm_convolve_s8()
cortex_m::quantized_linear_per_tensor_out — Fully connected layer using arm_fully_connected_s8()
cortex_m::quantized_avg_pool2d.out — Average pooling
cortex_m::quantized_max_pool2d.out — Max pooling
- Convolution (conv2d)
- Depthwise Convolution (depthwise_conv2d)
Activation Functions
cortex_m::quantized_relu.out — ReLU activation
cortex_m::quantized_hardtanh.out — Clamped activation function
-
E2E integration of running mobile vision models
-
Performance full fledged Analysis w.r.t TFLM and Benchmarking
Alternatives
No response
Additional context
No response
RFC (Optional)
No response
Sub-issues
Metadata
Metadata
Assignees
Labels
Type
Projects
Status