NN2: version 1.10

Makefile

@@ -12,12 +12,7 @@ endif
 export CROSS_COMPILE INSTALL_DIR
 
 
-all: nn2_ref_x86 nn2_ref
-
-nn2_ref:
-	DSP_LIB="libcsi_nn2_ref" CFLAGS="-mcpu=c860v -DCSI_BUILD_REF $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="csky-abiv2-linux-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_ref -j8
-	cd source/; find . -name *.o | xargs rm; cd -
+all: nn2_ref_x86
 
 nn2_c860:
 	DSP_LIB="libcsi_nn2_c860" CFLAGS="-mcpu=c860v -DCSI_BUILD_REF $(EXTRA_CFLAGS)" \
@@ -37,48 +32,28 @@ nn2_ref_x86:
 	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_ref nn2_shared -j8
 	cd source/; find . -name *.o | xargs rm; cd -
 
-nn2_openvx:
-	DSP_LIB="libcsi_nn2_openvx.a" CFLAGS="-mcpu=c860v -DCSI_BUILD_OPENVX -mhard-float $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="csky-abiv2-linux-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_openvx -j8
-	cd source/; find . -name *.o | xargs rm; cd -
-	DSP_LIB="libcsi_nn2_openvx.so" CFLAGS="-mcpu=c860v -fPIC -DCSI_BUILD_OPENVX -mhard-float $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="csky-abiv2-linux-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_openvx nn2_shared -j8
+nn2_ref_i805:
+	DSP_LIB="libcsi_nn2_ref_i805.a" CFLAGS="-DCSI_BUILD_REF_I805 -DCSI_MATH_DSP -mcpu=i805 $(EXTRA_CFLAGS)" \
+	CROSS_COMPILE="csky-abiv2-elf-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_ref_i805 -j8
 	cd source/; find . -name *.o | xargs rm; cd -
 
-nn2_pnna:
-	DSP_LIB="libcsi_nn2_pnna.a" CFLAGS="-DCSI_BUILD_PNNA $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="riscv64-unknown-linux-gnu-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_pnna -j8
-	cd source/; find . -name *.o | xargs rm; cd -
-	DSP_LIB="libcsi_nn2_pnna.so" CFLAGS="-fPIC -DCSI_BUILD_PNNA $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="riscv64-unknown-linux-gnu-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_pnna nn2_shared -j8
+nn2_e804:
+	DSP_LIB="libcsi_nn2_e804.a" CFLAGS="-DCSI_BUILD_E804 -mcpu=e804d -mno-required-attr-fpu-abi $(EXTRA_CFLAGS)" \
+	CROSS_COMPILE="csky-abiv2-elf-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_e804 -j8
 	cd source/; find . -name *.o | xargs rm; cd -
 
-nn2_pnna_x86:
-	DSP_LIB="libcsi_nn2_pnna_x86.a" CFLAGS="-DCSI_BUILD_PNNA $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_pnna -j8
-	cd source/; find . -name *.o | xargs rm; cd -
-	DSP_LIB="libcsi_nn2_pnna_x86.so" CFLAGS="-fPIC -DCSI_BUILD_PNNA $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_pnna nn2_shared_x86 -j8
+nn2_i805:
+	DSP_LIB="libcsi_nn2_i805.a" CFLAGS="-DCSI_BUILD_I805 -DCSI_BUILD_REF -DCSI_BUILD_GREF -mcpu=ck805ef -mhard-float $(EXTRA_CFLAGS)" \
+	CROSS_COMPILE="csky-abiv2-elf-" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_i805 -j8
 	cd source/; find . -name *.o | xargs rm; cd -
 
-nn2_gref:
-	DSP_LIB="libcsi_nn2_gref.a" CFLAGS="-DCSI_BUILD_REF -DCSI_BUILD_GREF $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_gref -j8
-	cd source/; find . -name *.o | xargs rm; cd -
-	DSP_LIB="libcsi_nn2_gref.so" CFLAGS="-fPIC -DCSI_BUILD_REF -DCSI_BUILD_GREF $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_gref nn2_shared -j8
-	cd source/; find . -name *.o | xargs rm; cd -
-
-nn2_dp1k:
-	DSP_LIB="libcsi_nn2_dp1000.so" CFLAGS="-fPIC -DCSI_BUILD_DP1K -DCSI_BUILD_REF $(EXTRA_CFLAGS)" \
-	CROSS_COMPILE="" NN2_ROOT=${NN2_ROOT} make -C build_script/nn2_dp1k nn2_shared -j8
-	cd source/; find . -name *.o | xargs rm; cd -
 
 .PHONY: install_nn2
 install_nn2: include
 	mkdir -p install_nn2/lib
 	cp include install_nn2 -r
 	cp lib/libcsi_nn2_* install_nn2/lib -rf
+	cp version install_nn2/ -rf
 
 
 clean:

README.md

@@ -4,15 +4,16 @@ CSI-NN2 是 T-HEAD 提供的一组针对无剑 SoC 平台的神经网络库 API
 
 CSI-NN2 的特性：
 
-- 开源 c 代码版本的参考实现。
-- 提供玄铁 CPU 的汇编优化实现。
-- 支持 i8/u8 数据类型的对称和非对称量化。
+- C 代码版本的参考实现。
+- 提供玄铁系列 CPU 的汇编优化实现。
+- 支持对称量化和非对称量化。
+- 支持8位定点，16位定点和16位浮点等数据类型。
 - 兼容 NCHW 和 NHWC 格式。
-- 搭配 HHB 实现代码自动调用。
-- 覆盖 CPU，GPU, NPU 体系结构。
-- 附加一些接口辅助使用，作为参考实现。
+- 搭配 [HHB](https://occ.t-head.cn/development/series/index?spm=a2cl5.14300690.0.0.4aca475a4yHCxV&id=3865005559921381376&type=kind) 实现代码自动调用。
+- 覆盖 CPU，NPU 架构。
+- 附加一些辅助接口，参考使用。
 
-CSI-NN2 只提供接口声明和接口的参考实现，对各个接口的优化工作交由各个设备提供商完成。
+CSI-NN2 提供了完成的接口声明和接口的参考实现，各个设备提供商可以依此针对性的完成各个接口的优化工作。
 
 ## 文档说明
 
@@ -25,3 +26,5 @@ CSI-NN2 参考、借鉴了下列项目：
 - [Tensorflow](https://github.com/tensorflow/tensorflow)
 - [ncnn](https://github.com/Tencent/ncnn)
 - [MNN](https://github.com/alibaba/MNN)
+- [Tengine](https://github.com/OAID/Tengine)
+- [CMSIS_5](https://github.com/ARM-software/CMSIS_5)

build_script/nn2.objs

@@ -157,4 +157,5 @@ objects += ${NN2_ROOT}/source/nn2/utils.o
 objects += ${NN2_ROOT}/source/nn2/setup.o
 objects += ${NN2_ROOT}/source/nn2/node.o
 objects += ${NN2_ROOT}/source/nn2/scatter.o
+objects += ${NN2_ROOT}/source/nn2/memory.o
 objects += ${NN2_ROOT}/source/nn2/debug.o

build_script/nn2_c906/nn2_c906.objs

@@ -1,22 +1,47 @@
 objects += ${NN2_ROOT}/source/c906_opt/abs.o
 
 objects += ${NN2_ROOT}/source/c906_opt/add.o
+objects += ${NN2_ROOT}/source/c906_opt/sub.o
+objects += ${NN2_ROOT}/source/c906_opt/mul.o
+objects += ${NN2_ROOT}/source/c906_opt/minimum.o
 
 objects += ${NN2_ROOT}/source/c906_opt/broadcast_to.o
 objects += ${NN2_ROOT}/source/c906_opt/clip.o
+objects += ${NN2_ROOT}/source/c906_opt/concat.o
+objects += ${NN2_ROOT}/source/c906_opt/split.o
 objects += ${NN2_ROOT}/source/c906_opt/convolution.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_relu.o
 
 objects += ${NN2_ROOT}/source/c906_opt/fullyconnected.o
 objects += ${NN2_ROOT}/source/c906_opt/maxpool.o
 objects += ${NN2_ROOT}/source/c906_opt/global_maxpool.o
 objects += ${NN2_ROOT}/source/c906_opt/avgpool.o
 objects += ${NN2_ROOT}/source/c906_opt/global_avgpool.o
 
+objects += ${NN2_ROOT}/source/c906_opt/pad.o
 objects += ${NN2_ROOT}/source/c906_opt/prelu.o
 objects += ${NN2_ROOT}/source/c906_opt/relu.o
 objects += ${NN2_ROOT}/source/c906_opt/relu1.o
 objects += ${NN2_ROOT}/source/c906_opt/relu6.o
 
 objects += ${NN2_ROOT}/source/c906_opt/leaky_relu.o
+objects += ${NN2_ROOT}/source/c906_opt/utils.o
 
 objects += ${NN2_ROOT}/source/c906_opt/setup.o
+
+objects += ${NN2_ROOT}/source/c906_opt/sgemm.o
+objects += ${NN2_ROOT}/source/c906_opt/gemm_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_1x1.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_1x1_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_3x3.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_3x3_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_sgemm.o
+objects += ${NN2_ROOT}/source/c906_opt/convolution_gemm_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_3x3.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_3x3_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_3x3_pack4.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_3x3_pack8_fp16.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_5x5.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_relu_3x3.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_relu_3x3_pack4.o
+objects += ${NN2_ROOT}/source/c906_opt/depthwise_convolution_relu_5x5.o

build_script/nn2_e804/Makefile

@@ -0,0 +1,21 @@
+CROSS_COMPILE   ?= csky-abiv2-elf-
+CC        	= $(CROSS_COMPILE)gcc
+AR        	= $(CROSS_COMPILE)ar
+
+CFLAGS += -ffunction-sections -fdata-sections -Wl,--gc-sections
+INCLUDE = -I../../include/include_xt800/dsp_include/ -I../../include/include_xt800/csi_core_include/ -I../../include/include_xt800/nn_include 
+INCLUDE +=  -I../../include
+
+all: nn2_e804
+
+include nn2_dsp2.objs
+include ../nn2.objs
+include ../nn2_ref/nn2.objs
+
+nn2_e804: $(objects)
+	$(AR) rcs $(INSTALL_DIR)/$(DSP_LIB) $^
+
+%.o: %.c
+	$(CC) -c $(CFLAGS) $(INCLUDE) $(MACRO) $(ENDIAN) $< -o $@
+%.o: %.S
+	$(CC) -c $(CFLAGS) $(INCLUDE) $(MACRO) $(ENDIAN) $< -o $@

build_script/nn2_e804/nn2_dsp2.objs

@@ -0,0 +1,38 @@
+#activation
+objects += ${NN2_ROOT}/source/e804_opt/activation/csi_xt800p_nn_activations_q15.o
+objects += ${NN2_ROOT}/source/e804_opt/activation/csi_xt800p_nn_activations_q7.o
+objects += ${NN2_ROOT}/source/e804_opt/activation/csi_xt800p_relu_q15.o
+objects += ${NN2_ROOT}/source/e804_opt/activation/csi_xt800p_relu_q7.o
+
+#convolution
+objects += ${NN2_ROOT}/source/e804_opt/convolution/csi_xt800p_convolve_1x1_HWC_q7_fast.o
+objects += ${NN2_ROOT}/source/e804_opt/convolution/csi_xt800p_convolve_HWC_q15_basic.o
+objects += ${NN2_ROOT}/source/e804_opt/convolution/csi_xt800p_convolve_HWC_q7_RGB.o
+objects += ${NN2_ROOT}/source/e804_opt/convolution/csi_xt800p_convolve_HWC_q7_basic.o
+objects += ${NN2_ROOT}/source/e804_opt/convolution/csi_xt800p_depthwise_separable_conv_HWC_q7.o
+
+#fully-connect
+objects += ${NN2_ROOT}/source/e804_opt/fully-connect/csi_xt800p_fully_connected_mat_q7_vec_q15.o
+objects += ${NN2_ROOT}/source/e804_opt/fully-connect/csi_xt800p_fully_connected_q15.o
+objects += ${NN2_ROOT}/source/e804_opt/fully-connect/csi_xt800p_fully_connected_q7.o
+
+#nn-support
+objects += ${NN2_ROOT}/source/e804_opt/nn-support/csi_xt800p_nntables.o
+
+#pooling
+objects += ${NN2_ROOT}/source/e804_opt/pooling/csi_xt800p_pool_q7_HWC.o
+
+#softmax
+objects += ${NN2_ROOT}/source/e804_opt/softmax/csi_xt800p_softmax_q15.o
+objects += ${NN2_ROOT}/source/e804_opt/softmax/csi_xt800p_softmax_q7.o
+
+objects += ${NN2_ROOT}/source/e804_opt/convolution.o
+objects += ${NN2_ROOT}/source/e804_opt/fullyconnected.o
+objects += ${NN2_ROOT}/source/e804_opt/avgpool.o
+objects += ${NN2_ROOT}/source/e804_opt/maxpool.o
+objects += ${NN2_ROOT}/source/e804_opt/softmax.o
+objects += ${NN2_ROOT}/source/e804_opt/relu.o
+objects += ${NN2_ROOT}/source/e804_opt/sigmoid.o
+objects += ${NN2_ROOT}/source/e804_opt/tanh.o
+
+objects += ${NN2_ROOT}/source/e804_opt/setup.o

build_script/nn2_gref/nn2.objs

@@ -152,3 +152,4 @@ objects += ${NN2_ROOT}/source/graph_ref/reorg.o
 objects += ${NN2_ROOT}/source/graph_ref/sequence_mask.o
 objects += ${NN2_ROOT}/source/graph_ref/where.o
 objects += ${NN2_ROOT}/source/graph_ref/space_to_batch_nd.o
+objects += ${NN2_ROOT}/source/graph_ref/subgraph.o

build_script/nn2_i805/Makefile

@@ -0,0 +1,22 @@
+CROSS_COMPILE   ?= csky-abiv2-elf-
+CC        	= $(CROSS_COMPILE)gcc
+AR        	= $(CROSS_COMPILE)ar
+
+CFLAGS += -ffunction-sections -fdata-sections -Wl,--gc-sections
+INCLUDE = -I../../include/include_xt800/dsp_include/ -I../../include/include_xt800/csi_core_include/ -I../../include/include_xt800/nn_include
+INCLUDE +=  -I../../include
+
+all: nn2_i805
+
+include nn2_vdsp2.objs
+include ../nn2.objs
+include ../nn2_ref/nn2.objs
+include ../nn2_gref/nn2.objs
+
+nn2_i805: $(objects)
+	$(AR) rcs $(INSTALL_DIR)/$(DSP_LIB) $^
+
+%.o: %.c
+	$(CC) -c $(CFLAGS) $(INCLUDE) $(MACRO) $(ENDIAN) $< -o $@
+%.o: %.S
+	$(CC) -c $(CFLAGS) $(INCLUDE) $(MACRO) $(ENDIAN) $< -o $@

build_script/nn2_i805/nn2_vdsp2.objs

@@ -0,0 +1,72 @@
+#activation
+#objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_nn_activations_q15.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_nn_activations_q15_fast.o
+#objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_nn_activations_q7.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_nn_activations_q7_fast.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_relu_q15.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_xt800v_relu_q7.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_i805_relu_8.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_i805_relu6_8.o
+objects += ${NN2_ROOT}/source/i805_opt/activation/csi_i805_clip_8.o
+
+#convolution
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_convolve_1x1_HWC_q7_fast.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_convolve_HWC_q15_basic.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_convolve_HWC_q7_RGB.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_convolve_HWC_q7_basic.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_convolve_HWC_q7_fast_nonsquare.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_depthwise_separable_conv_HWC_q7.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_xt800v_depthwise_separable_conv_HWC_q7_nonsquare.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_i805_convolution_1x1_8.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_i805_convolution_8.o
+objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_i805_depthwise_convolution_8.o
+#objects += ${NN2_ROOT}/source/i805_opt/convolution/csi_i805_depthwise_convolution_3x3_8.o
+
+
+#fully-connect
+objects += ${NN2_ROOT}/source/i805_opt/fully-connect/csi_xt800v_fully_connected_mat_q7_vec_q15.o
+objects += ${NN2_ROOT}/source/i805_opt/fully-connect/csi_xt800v_fully_connected_q15.o
+objects += ${NN2_ROOT}/source/i805_opt/fully-connect/csi_xt800v_fully_connected_q7x4.o
+#objects += ${NN2_ROOT}/source/i805_opt/fully-connect/csi_xt800v_fully_connected_q7x16.o
+objects += ${NN2_ROOT}/source/i805_opt/fully-connect/csi_i805_fullyconnected_8.o
+
+#nn-support
+objects += ${NN2_ROOT}/source/i805_opt/nn-support/csi_xt800v_nntables.o
+
+#pooling
+objects += ${NN2_ROOT}/source/i805_opt/pooling/csi_xt800v_pool_q7_HWC.o
+objects += ${NN2_ROOT}/source/i805_opt/pooling/csi_xt800v_avepool_q7_HWC_nonsquare.o
+objects += ${NN2_ROOT}/source/i805_opt/pooling/csi_i805_maxpool_8.o
+
+#softmax
+objects += ${NN2_ROOT}/source/i805_opt/softmax/csi_xt800v_softmax_q15.o
+objects += ${NN2_ROOT}/source/i805_opt/softmax/csi_xt800v_softmax_q7.o
+#objects += ${NN2_ROOT}/source/i805_opt/softmax/csi_i805_softmax_8.o
+
+#gemm
+#objects += ${NN2_ROOT}/source/i805_opt/gemm/csi_i805_gemm_4x4_8.o
+#objects += ${NN2_ROOT}/source/i805_opt/gemm/csi_i805_gemm_4x16_8.o
+objects += ${NN2_ROOT}/source/i805_opt/gemm/csi_i805_vec_mat_mult_8.o
+objects += ${NN2_ROOT}/source/i805_opt/gemm/csi_i805_mat_mult_nt_t_8.o
+
+#basic_math_func
+objects += ${NN2_ROOT}/source/i805_opt/basic_math/csi_i805_elementwise_add_8.o
+objects += ${NN2_ROOT}/source/i805_opt/basic_math/csi_i805_elementwise_mul_8.o
+
+#reshape
+objects += ${NN2_ROOT}/source/i805_opt/reshape/csi_i805_reshape_8.o
+
+objects += ${NN2_ROOT}/source/i805_opt/convolution.o
+objects += ${NN2_ROOT}/source/i805_opt/fullyconnected.o
+objects += ${NN2_ROOT}/source/i805_opt/add.o
+objects += ${NN2_ROOT}/source/i805_opt/avgpool.o
+objects += ${NN2_ROOT}/source/i805_opt/maxpool.o
+objects += ${NN2_ROOT}/source/i805_opt/mul.o
+objects += ${NN2_ROOT}/source/i805_opt/softmax.o
+objects += ${NN2_ROOT}/source/i805_opt/relu.o
+objects += ${NN2_ROOT}/source/i805_opt/relu6.o
+objects += ${NN2_ROOT}/source/i805_opt/clip.o
+objects += ${NN2_ROOT}/source/i805_opt/reshape.o
+objects += ${NN2_ROOT}/source/i805_opt/sigmoid.o
+objects += ${NN2_ROOT}/source/i805_opt/tanh.o
+objects += ${NN2_ROOT}/source/i805_opt/setup.o