Merge pull request #91 from tum-ei-eda/fix-integration

Fix integration

.github/workflows/integration_tests.yml

@@ -29,7 +29,7 @@ jobs:
         matrix:
           tflm_ref:
           - main  # latest upstream
-          - 19aaea85e4679a9a2f265e07ba190ac5ea4d3766 # Last working commit (needs to be updated manually)
+          - 8eb6b23de4470d6a8da3131650d6a67514dfa130  # Last working commit (needs to be updated manually)
           config:
             - {vext: "OFF", portable: "OFF", pext: "OFF", mode: "release_with_logs", toolchain: "gcc", arch: "native", vlen: 0, elen: 64, sim: "Native"}
             - {vext: "OFF", portable: "OFF", pext: "OFF", mode: "release_with_logs", toolchain: "gcc", arch: "rv32gc", vlen: 0, elen: 64, sim: "Spike"}
@@ -227,7 +227,7 @@ jobs:
         matrix:
           tflm_ref:
           - main  # latest upstream
-          - 19aaea85e4679a9a2f265e07ba190ac5ea4d3766 # Last working commit (needs to be updated manually)
+          - 8eb6b23de4470d6a8da3131650d6a67514dfa130  # Last working commit (needs to be updated manually)
           config:
             - {vext_flag: "", portable_flag: "", pext_flag: "", toolchain: "x86", arch: "", vlen: 0, elen: 0, sim: "Native"}
             - {vext_flag: "", portable_flag: "", pext_flag: "", toolchain: "GCC", arch: "rv32gc", vlen: 0, elen: 0, sim: "Spike"}

Include/CMSIS/NN/Include/arm_nn_types.h

@@ -32,13 +32,15 @@ extern "C" {
 #define cmsis_nn_bias_data muriscv_nn_bias_data
 #define cmsis_nn_per_channel_quant_params muriscv_nn_per_channel_quant_params
 #define cmsis_nn_per_tensor_quant_params muriscv_nn_per_tensor_quant_params
+#define cmsis_nn_quant_params muriscv_nn_quant_params
 #define cmsis_nn_activation muriscv_nn_activation
 #define cmsis_nn_conv_params muriscv_nn_conv_params
 #define cmsis_nn_dw_conv_params muriscv_nn_dw_conv_params
 #define cmsis_nn_pool_params muriscv_nn_pool_params
 #define cmsis_nn_fc_params muriscv_nn_fc_params
 #define cmsis_nn_svdf_params muriscv_nn_svdf_params
 #define cmsis_nn_softmax_lut_s16 muriscv_nn_softmax_lut_s16
+#define cmsis_nn_bmm_params muriscv_nn_bmm_params
 
 #define cmsis_nn_lstm_params muriscv_nn_lstm_params
 #define cmsis_nn_lstm_context muriscv_nn_lstm_context

Include/CMSIS/NN/Include/arm_nnfunctions.h

@@ -90,6 +90,8 @@ extern "C" {
 #define arm_depthwise_conv_s4_opt_get_buffer_size muriscv_nn_depthwise_conv_s4_opt_get_buffer_size
 #define arm_fully_connected_s4 muriscv_nn_fully_connected_s4
 #define arm_fully_connected_s8 muriscv_nn_fully_connected_s8
+#define arm_fully_connected_per_channel_s8 muriscv_nn_fully_connected_per_channel_s8
+#define arm_fully_connected_wrapper_s8 muriscv_nn_fully_connected_wrapper_s8
 #define arm_vector_sum_s8 muriscv_nn_vector_sum_s8
 #define arm_vector_sum_s8_s64 muriscv_nn_vector_sum_s8_s64
 #define arm_fully_connected_s8_get_buffer_size muriscv_nn_fully_connected_s8_get_buffer_size
@@ -122,6 +124,7 @@ extern "C" {
 #define arm_softmax_s16 muriscv_nn_softmax_s16
 #define arm_softmax_u8 muriscv_nn_softmax_u8
 #define arm_reshape_s8 muriscv_nn_reshape_s8
+#define arm_transpose_s8 muriscv_nn_transpose_s8
 #define arm_concatenation_s8_x muriscv_nn_concatenation_s8_x
 #define arm_concatenation_s8_y muriscv_nn_concatenation_s8_y
 #define arm_concatenation_s8_z muriscv_nn_concatenation_s8_z
@@ -133,6 +136,11 @@ extern "C" {
 #define arm_svdf_s8_get_buffer_size_mve muriscv_nn_svdf_s8_get_buffer_size_mve
 #define arm_lstm_unidirectional_s8 muriscv_nn_lstm_unidirectional_s8
 #define arm_lstm_unidirectional_s16 muriscv_nn_lstm_unidirectional_s16
+#define arm_batch_matmul_s8 muriscv_nn_batch_matmul_s8
+#define arm_batch_matmul_s16 muriscv_nn_batch_matmul_s16
+#define arm_pad_s8 muriscv_nn_pad_s8
+#define arm_minimum_s8 muriscv_nn_minimum_s8
+#define arm_maximum_s8 muriscv_nn_maximum_s8
 #ifdef __cplusplus
 }
 #endif

Include/muriscv_nn_support_functions.h

@@ -22,8 +22,8 @@
  * Title:        muriscv_nn_support_functions.h
  * Description:  Public header file of support functions for MURISCV NN Library
  *
- * $Date:        27 May 2024
- * $Revision:    V.22.1.0
+ * $Date:        08 October 2024
+ * $Revision:    V.22.4.0
  *
  * Target :  Arm(R) M-Profile Architecture
  * -------------------------------------------------------------------- */
@@ -849,7 +849,48 @@ muriscv_nn_status muriscv_nn_vec_mat_mult_t_s8(const int8_t *lhs,
                                              const int32_t rhs_offset);
 
 /**
- * @brief s16 Vector by Matrix (transposed) multiplication
+ * @brief s8 Vector by Matrix (transposed) multiplication using per channel quantization for output
+ *
+ * @param[in]      lhs             Input left-hand side vector
+ * @param[in]      rhs             Input right-hand side matrix (transposed)
+ * @param[in]      kernel_sum      Kernel sums of the kernels (rhs). See muriscv_nn_vector_sum_s8 for more info.
+ * @param[in]      bias            Input bias
+ * @param[out]     dst             Output vector
+ * @param[in]      lhs_offset      Offset to be added to the input values of the left-hand side vector.
+ *                                 Range: -127 to 128
+ * @param[in]      dst_offset      Offset to be added to the output values. Range: -127 to 128
+ * @param[in]      dst_multiplier  Output multipliers
+ * @param[in]      dst_shift       Output shifts
+ * @param[in]      rhs_cols        Number of columns in the right-hand side input matrix
+ * @param[in]      rhs_rows        Number of rows in the right-hand side input matrix
+ * @param[in]      activation_min  Minimum value to clamp the output to. Range: int8
+ * @param[in]      activation_max  Maximum value to clamp the output to. Range: int8
+ * @param[in]      address_offset  Memory position offset for dst. First output is stored at 'dst', the
+ *                                 second at 'dst + address_offset' and so on. Default value is typically 1.
+ * @param[in]      rhs_offset      Offset to be added to the input values of the right-hand side vector.
+ *                                 Range: -127 to 128
+ *
+ * @return         The function returns <code>MURISCV_NN_SUCCESS</code>
+ *
+ */
+muriscv_nn_status muriscv_nn_vec_mat_mult_t_per_ch_s8(const int8_t *lhs,
+                                                    const int8_t *rhs,
+                                                    const int32_t *kernel_sum,
+                                                    const int32_t *bias,
+                                                    int8_t *dst,
+                                                    const int32_t lhs_offset,
+                                                    const int32_t dst_offset,
+                                                    const int32_t *dst_multiplier,
+                                                    const int32_t *dst_shift,
+                                                    const int32_t rhs_cols,
+                                                    const int32_t rhs_rows,
+                                                    const int32_t activation_min,
+                                                    const int32_t activation_max,
+                                                    const int32_t address_offset,
+                                                    const int32_t rhs_offset);
+
+/**
+ * @brief s16 Vector by s8 Matrix (transposed) multiplication
  *
  * @param[in]      lhs             Input left-hand side vector
  * @param[in]      rhs             Input right-hand side matrix (transposed)
@@ -876,6 +917,34 @@ muriscv_nn_status muriscv_nn_vec_mat_mult_t_s16(const int16_t *lhs,
                                               const int32_t activation_min,
                                               const int32_t activation_max);
 
+/**
+ * @brief s16 Vector by s16 Matrix (transposed) multiplication
+ *
+ * @param[in]      lhs             Input left-hand side vector
+ * @param[in]      rhs             Input right-hand side matrix (transposed)
+ * @param[in]      bias            Input bias
+ * @param[out]     dst             Output vector
+ * @param[in]      dst_multiplier  Output multiplier
+ * @param[in]      dst_shift       Output shift
+ * @param[in]      rhs_cols        Number of columns in the right-hand side input matrix
+ * @param[in]      rhs_rows        Number of rows in the right-hand side input matrix
+ * @param[in]      activation_min  Minimum value to clamp the output to. Range: int16
+ * @param[in]      activation_max  Maximum value to clamp the output to. Range: int16
+ *
+ * @return         The function returns <code>MURISCV_NN_SUCCESS</code>
+ *
+ */
+muriscv_nn_status muriscv_nn_vec_mat_mult_t_s16_s16(const int16_t *lhs,
+                                                  const int16_t *rhs,
+                                                  const int64_t *bias,
+                                                  int16_t *dst,
+                                                  const int32_t dst_multiplier,
+                                                  const int32_t dst_shift,
+                                                  const int32_t rhs_cols,
+                                                  const int32_t rhs_rows,
+                                                  const int32_t activation_min,
+                                                  const int32_t activation_max);
+
 /**
  * @brief s8 Vector by Matrix (transposed) multiplication with s16 output
  *
@@ -2488,6 +2557,26 @@ muriscv_nn_status muriscv_nn_elementwise_mul_acc_s16(const int16_t *input_1_vect
                                                 const int32_t out_activation_max,
                                                 const int32_t block_size);
 
+/**
+ * @brief Check if a broadcast is required between 2 muriscv_nn_dims.
+ * @param[in]       shape_1             pointer to input tensor 1
+ * @param[in]       shape_2             pointer to input tensor 2
+ * @return          The function returns 1 if a broadcast is required, or 0 if not.
+ *
+ * @details   Compares each dimension and returns 1 if any dimension does not match.
+ *            This function does not check that broadcast rules are met.
+ */
+__STATIC_FORCEINLINE int32_t muriscv_nn_check_broadcast_required(const muriscv_nn_dims *shape_1, const muriscv_nn_dims *shape_2)
+{
+    if ((shape_1->n != shape_2->n) || (shape_1->h != shape_2->h) || (shape_1->w != shape_2->w) ||
+        (shape_1->c != shape_2->c))
+    {
+        return 1;
+    }
+
+    return 0;
+}
+
 #ifdef __cplusplus
 }
 #endif

Include/muriscv_nn_types.h

@@ -89,6 +89,17 @@ typedef struct
     int32_t shift;      /**< Shift value */
 } muriscv_nn_per_tensor_quant_params;
 
+/** MURISCV-NN object for quantization parameters.
+ *  This struct supports both per-tensor and per-channels requantization
+ *  and is recommended for new operators.
+ */
+typedef struct
+{
+    int32_t *multiplier;    /**< Multiplier values */
+    int32_t *shift;         /**< Shift values */
+    int32_t is_per_channel; /** Indicating if per channel or per tensor quantization */
+} muriscv_nn_quant_params;
+
 /** MURISCV-NN object for the quantized Relu activation */
 typedef struct
 {
@@ -148,6 +159,21 @@ typedef struct
     muriscv_nn_activation activation;
 } muriscv_nn_fc_params;
 
+/** MURISCV-NN object for Batch Matmul layer parameters */
+typedef struct
+{
+    const bool adj_x;
+    const bool adj_y;
+    muriscv_nn_fc_params fc_params;
+} muriscv_nn_bmm_params;
+
+/** MURISCV-NN object for Transpose layer parameters */
+typedef struct
+{
+    const int32_t num_dims;
+    const uint32_t *permutations; /**< The dimensions applied to the input dimensions */
+} muriscv_nn_transpose_params;
+
 /** MURISCV-NN object for SVDF layer parameters */
 typedef struct
 {

Integration/tflm/CMakeLists.txt

@@ -271,8 +271,8 @@ set(TFLM_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/micro/micro_log.cc
   ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/micro/tflite_bridge/flatbuffer_conversions_bridge.h
   ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/micro/tflite_bridge/flatbuffer_conversions_bridge.cc
-  ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/core/api/error_reporter.cc
-  ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/core/api/error_reporter.h
+  ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/compiler/mlir/lite/core/api/error_reporter.cc
+  ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/compiler/mlir/lite/core/api/error_reporter.h
 #  ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/core/api/op_resolver.cc
   ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/core/api/flatbuffer_conversions.cc
   ${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite/core/api/tensor_utils.h

Integration/tflm/download_tflm.sh

@@ -23,6 +23,8 @@ set -euo pipefail
 # Path to this script
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
+TFLM_REF=${TFLM_REF:-""}
+
 echo "Download TFLM sources."
 if [ ! -d "tflite-micro" ]; then
   git clone https://github.com/tensorflow/tflite-micro.git

Source/BasicMathFunctions/CMakeLists.txt

@@ -24,4 +24,6 @@ target_sources(${MURISCVNN_LIB} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_e
                                         ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_elementwise_mul_s16.c
                                         ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_elementwise_mul_s16_s8.c
                                         ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_elementwise_mul_acc_s16.c
-                                        ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_elementwise_mul_s16_batch_offset.c)
+                                        ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_elementwise_mul_s16_batch_offset.c
+                                        ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_maximum_s8.c
+                                        ${CMAKE_CURRENT_SOURCE_DIR}/muriscv_nn_minimum_s8.c)