Star1280
2051
代码
代码
Issues586
Pull Requests580
Actions
项目讨论
Wiki
分析
Star1280
2051
已合并
docs: 统一原型注释中的产品名称（9.2.0） #11179
yuning_chen创建于 9月24日
docs: 统一原型注释中的产品名称（9.2.0） #11179
已合并
yuning_chen创建于 9月24日
从yolic/ops-nnsoc920合入到cann/ops-nn9.2.0
共 70 个文件变更+254-282
Mactivation/fast_gelu/op_graph/fast_gelu_proto.h+3-3
@@ -17,8 +17,8 @@ namespace ge {
17 17 
18/**18/**
19* @brief19* @brief
20-* For Atlas 200/300/500 Inference Product, Atlas Training Series Product,20+* For Atlas training products,
21- Atlas Inference Series Product, Ascend 610 AI Processor,21+ Atlas inference products, Ascend 610 AI Processor,
22 the calculation formula is x*e^(0.851*x)*(x-|x|)/(1+e^(-1.702|x|)).22 the calculation formula is x*e^(0.851*x)*(x-|x|)/(1+e^(-1.702|x|)).
23* For other chips, the calculation formula is x/(1+e^(-1.702*x)).23* For other chips, the calculation formula is x/(1+e^(-1.702*x)).
24 24 
@@ -37,4 +37,4 @@ REG_OP(FastGelu)
37 .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))37 .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
38 .OP_END_FACTORY_REG(FastGelu)38 .OP_END_FACTORY_REG(FastGelu)
39} // namespace ge39} // namespace ge
40-#endif // OPS_BUILT_IN_OP_PROTO_INC_NONLINEAR_FUC_OPS_H_40+#endif // OPS_BUILT_IN_OP_PROTO_INC_NONLINEAR_FUC_OPS_H_
Mactivation/ge_glu_grad_v2/op_graph/ge_glu_grad_v2_proto.h+2-2
@@ -40,7 +40,7 @@ namespace ge {
40* @li dim: An optional Int. The dimension to be split, default is -1.40* @li dim: An optional Int. The dimension to be split, default is -1.
41* @li approximate: An optional Int. Determines which formula to use for the activation computation.41* @li approximate: An optional Int. Determines which formula to use for the activation computation.
42* The gelu grad approximation algorithm to use: 0('none') or 1('tanh'), default is 1('tanh').42* The gelu grad approximation algorithm to use: 0('none') or 1('tanh'), default is 1('tanh').
43-* Atlas Inference Series Product only supports 'tanh'(1).43+* Atlas inference products only supports 'tanh'(1).
44* @li activate_left: An optional Bool.44* @li activate_left: An optional Bool.
45* Whether the left side of x is used as an input parameter to the activation function,45* Whether the left side of x is used as an input parameter to the activation function,
46* default is false, use the right side.46* default is false, use the right side.
@@ -62,4 +62,4 @@ REG_OP(GeGluGradV2)
62 62 
63} // namespace ge63} // namespace ge
64 64 
65-#endif65+#endif
Mactivation/ge_glu_v2/op_graph/ge_glu_v2_proto.h+2-2
@@ -43,7 +43,7 @@ namespace ge {
43* @li dim: An optional int. The dimension to be split, default is -1.43* @li dim: An optional int. The dimension to be split, default is -1.
44* @li approximate: An optional int. Which formula used for the activation computation.44* @li approximate: An optional int. Which formula used for the activation computation.
45* The gelu approximation algorithm to use: 'none'(0) or 'tanh'(1), default is 'tanh'(1).45* The gelu approximation algorithm to use: 'none'(0) or 'tanh'(1), default is 'tanh'(1).
46-* Atlas Inference Series Product only support 'tanh'(1).46+* Atlas inference products only support 'tanh'(1).
47* @li activate_left: An optional bool.47* @li activate_left: An optional bool.
48* The gelu activate_left algorithm to use:48* The gelu activate_left algorithm to use:
49* 'false'(activate right) or 'true'(activate left), defalut is 'false'(activate right).49* 'false'(activate right) or 'true'(activate left), defalut is 'false'(activate right).
@@ -60,4 +60,4 @@ REG_OP(GeGluV2)
60 60 
61} // namespace ge61} // namespace ge
62 62 
63-#endif63+#endif
Mactivation/gelu_grad/op_graph/gelu_grad_proto.h+2-2
@@ -33,7 +33,7 @@ namespace ge {
33* @par Third-party framework compatibility33* @par Third-party framework compatibility
34* Compatible with the TensorFlow operator GeluGrad.34* Compatible with the TensorFlow operator GeluGrad.
35* @attention Constraints:35* @attention Constraints:
36-* In Ascend 950 AI Processor, inputs and outputs support broadcasting.36+* In Ascend 950PR&950DT products, inputs and outputs support broadcasting.
37*/37*/
38REG_OP(GeluGrad)38REG_OP(GeluGrad)
39 .INPUT(dy, TensorType({DT_BF16, DT_FLOAT16, DT_FLOAT}))39 .INPUT(dy, TensorType({DT_BF16, DT_FLOAT16, DT_FLOAT}))
@@ -44,4 +44,4 @@ REG_OP(GeluGrad)
44 44 
45} // namespace ge45} // namespace ge
46 46 
47-#endif47+#endif
Mactivation/log_softmax_grad/op_graph/log_softmax_grad_proto.h+1-1
@@ -27,7 +27,7 @@ namespace ge {
27 27 
28*@par Attributes:28*@par Attributes:
29* axis: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}".29* axis: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}".
30-* In Ascend 950 AI Processor, only single-axis reduction is supported. \n30+* In Ascend 950PR&950DT products, only single-axis reduction is supported. \n
31 31 
32*@par Outputs:32*@par Outputs:
33* y: A ND tensor. Has the same data type and shape as "grad". \n33* y: A ND tensor. Has the same data type and shape as "grad". \n
Mactivation/log_softmax_v2/op_graph/log_softmax_v2_proto.h+1-1
@@ -27,7 +27,7 @@ namespace ge {
27 27 
28*@par Attributes:28*@par Attributes:
29* axes: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}" .29* axes: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}" .
30-* In npuArch 3510 AI Processor, only single-axis reduction is supported. \n30+* In Ascend 950PR&950DT products, only single-axis reduction is supported. \n
31 31 
32*@par Outputs:32*@par Outputs:
33* logsoftmax: A ND tensor. Has the same data type as "logits" . \n33* logsoftmax: A ND tensor. Has the same data type as "logits" . \n
Mactivation/relu_grad_v2/op_graph/relu_grad_v2_proto.h+3-3
@@ -24,10 +24,10 @@ namespace ge {
24 * @par Inputs:24 * @par Inputs:
25 * Two inputs, including:25 * Two inputs, including:
26 * @li gradients: A tensor of input gradient. Indicates the gradient of backpropagation. Supported type:26 * @li gradients: A tensor of input gradient. Indicates the gradient of backpropagation. Supported type:
27- * TensorType::RealNumberType(). \n In Ascend 950 AI Processor, support ND format. Others support 5D, format must be27+ * TensorType::RealNumberType(). \n In Ascend 950PR&950DT products, support ND format. Others support 5D, format must be
28 * NC1HWC0.28 * NC1HWC0.
29 * @li mask: A tensor of input mask. Indicates the positive output. Must be the following types: uint8, uint1. \n29 * @li mask: A tensor of input mask. Indicates the positive output. Must be the following types: uint8, uint1. \n
30- * In Ascend 950 AI Processor, support ND format. Others support 5D, format must be NC1HWC0.30+ * In Ascend 950PR&950DT products, support ND format. Others support 5D, format must be NC1HWC0.
31 *31 *
32 * @par Outputs:32 * @par Outputs:
33 * backprops: A tensor of output. Indicates the backpropagation result.33 * backprops: A tensor of output. Indicates the backpropagation result.
@@ -46,4 +46,4 @@ REG_OP(ReluGradV2)
46 46 
47} // namespace ge47} // namespace ge
48 48 
49-#endif // OPS_BUILT_IN_OP_PROTO_INC_NONLINEAR_FUC_OPS_H_49+#endif // OPS_BUILT_IN_OP_PROTO_INC_NONLINEAR_FUC_OPS_H_
Mactivation/relu_v2/op_graph/relu_v2_proto.h+2-2
@@ -23,7 +23,7 @@ namespace ge {
23 *23 *
24 * @par Inputs:24 * @par Inputs:
25 * x: A tensor. Must be one of the following types: float32, float16, double, int8, int32, int16, int64, uint8, uint16,25 * x: A tensor. Must be one of the following types: float32, float16, double, int8, int32, int16, int64, uint8, uint16,
26- * qint8, bfloat16.\n In Ascend 950 AI Processor, support ND format. Others support 4D, format must be one of [NHWC,26+ * qint8, bfloat16.\n In Ascend 950PR&950DT products, support ND format. Others support 4D, format must be one of [NHWC,
27 * NCHW, HWCN].27 * NCHW, HWCN].
28 *28 *
29 * @par Outputs:29 * @par Outputs:
@@ -44,4 +44,4 @@ REG_OP(ReluV2)
44 .OUTPUT(mask, TensorType({DT_UINT8, DT_UINT1}))44 .OUTPUT(mask, TensorType({DT_UINT8, DT_UINT1}))
45 .OP_END_FACTORY_REG(ReluV2)45 .OP_END_FACTORY_REG(ReluV2)
46} // namespace ge46} // namespace ge
47-#endif47+#endif
Mactivation/softmax_grad/op_graph/softmax_grad_proto.h+2-2
@@ -29,7 +29,7 @@ namespace ge {
29 29 
30*@par Attributes:30*@par Attributes:
31* axes: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}".31* axes: An optional list of ints. Multi-axis reduction is supported. Defaults to "{-1}".
32-* In Ascend 950 AI Processor, only single-axis reduction is supported. \n32+* In Ascend 950PR&950DT products, only single-axis reduction is supported. \n
33 33 
34*@par Outputs:34*@par Outputs:
35*grad_x: A ND tensor. Has the same shape and data type as "softmax" . \n35*grad_x: A ND tensor. Has the same shape and data type as "softmax" . \n
@@ -44,4 +44,4 @@ REG_OP(SoftmaxGrad)
44 .ATTR(axes, ListInt, {-1})44 .ATTR(axes, ListInt, {-1})
45 .OP_END_FACTORY_REG(SoftmaxGrad)45 .OP_END_FACTORY_REG(SoftmaxGrad)
46} // namespace ge46} // namespace ge
47-#endif47+#endif
Mactivation/softmax_grad_ext/op_graph/softmax_grad_ext_proto.h+2-2
@@ -25,11 +25,11 @@ namespace ge {
25* @li grad: A tensor dtype of bfloat16, float16, float32.25* @li grad: A tensor dtype of bfloat16, float16, float32.
26* Indicates the reverse input tensor.26* Indicates the reverse input tensor.
27* The format must be FRACTAL_NZ unless specified. Support 2D ~ 6D.27* The format must be FRACTAL_NZ unless specified. Support 2D ~ 6D.
28-* In Ascend 950 AI Processor, the format must be ND.28+* In Ascend 950PR&950DT products, the format must be ND.
29 29 
30* @li x1: A tensor dtype of bfloat16, float16, float32. Indicates the inverse gradient.30* @li x1: A tensor dtype of bfloat16, float16, float32. Indicates the inverse gradient.
31* The format must be FRACTAL_NZ unless specified, Support 2D ~ 6D.31* The format must be FRACTAL_NZ unless specified, Support 2D ~ 6D.
32-* In Ascend 950 AI Processor, the format must be ND.32+* In Ascend 950PR&950DT products, the format must be ND.
33* Has the same type, format and shape as input grad.33* Has the same type, format and shape as input grad.
34* @li x2: A scalar dtype of bfloat16, float16, float32. Indicates the inverse scaling factor.34* @li x2: A scalar dtype of bfloat16, float16, float32. Indicates the inverse scaling factor.
35* The format must be ND. Has the same type as input grad. \n35* The format must be ND. Has the same type as input grad. \n
Mactivation/softmax_v2/op_graph/softmax_v2_proto.h+2-2
@@ -33,7 +33,7 @@ data types: float16, float32, bfloat16,
33* @par Attributes:33* @par Attributes:
34* @li axes: An optional list of int. Specifies on which dimensions of input x the Softmax operation is performed.34* @li axes: An optional list of int. Specifies on which dimensions of input x the Softmax operation is performed.
35* Multi-axis reduction is supported. Defaults to "{-1}".35* Multi-axis reduction is supported. Defaults to "{-1}".
36-* In Ascend 950 AI Processor, only single-axis reduction is supported. \n36+* In Ascend 950PR&950DT products, only single-axis reduction is supported. \n
37* @li half_to_float: An optional bool.37* @li half_to_float: An optional bool.
38* This parameter determines whether to convert the output data type to float32 when the input data type is float16.38* This parameter determines whether to convert the output data type to float32 when the input data type is float16.
39* Defaults to "false".39* Defaults to "false".
@@ -57,4 +57,4 @@ REG_OP(SoftmaxV2)
57 .OP_END_FACTORY_REG(SoftmaxV2)57 .OP_END_FACTORY_REG(SoftmaxV2)
58 58 
59} // namespace ge59} // namespace ge
60-#endif60+#endif
Mcommon/inc/op_graph/op_nn_proto_extend.h+31-34
@@ -933,7 +933,7 @@ following types: float16.
933* The data is stored in the order of: [out_channels].933* The data is stored in the order of: [out_channels].
934* @li offset_w: An optional quantitative offset tensor. Reserved.934* @li offset_w: An optional quantitative offset tensor. Reserved.
935*\n935*\n
936-* The following are the supported data types and data formats (except IPV350 and Ascend 950 AI Processor):936+* The following are the supported data types and data formats (except IPV350 and Ascend 950PR&950DT products):
937*\n937*\n
938| Tensor | x | filter | bias | y |\n938| Tensor | x | filter | bias | y |\n
939| :-------: | :------: | :------: | :------: | :------: |\n939| :-------: | :------: | :------: | :------: | :------: |\n
@@ -956,7 +956,7 @@ following types: float16.
956| Format | NCHW | NCHW | ND | NCHW |\n956| Format | NCHW | NCHW | ND | NCHW |\n
957| | NHWC | HWCN | ND | NHWC |\n957| | NHWC | HWCN | ND | NHWC |\n
958*\n958*\n
959-* The following are the supported data types and data formats for Ascend 950 AI Processor:959+* The following are the supported data types and data formats for Ascend 950PR&950DT products:
960*\n960*\n
961| Tensor | x | filter | bias | y |\n961| Tensor | x | filter | bias | y |\n
962| :-------: | :------: | :------: | :------: | :------: |\n962| :-------: | :------: | :------: | :------: | :------: |\n
@@ -1026,7 +1026,7 @@ following types: float16.
1026* int32(2147483647), the product of each dimension of x/filter/bias/offset_w/y1026* int32(2147483647), the product of each dimension of x/filter/bias/offset_w/y
1027* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x1027* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x
1028* exceeds the range in the above table, the correctness of the operator cannot be guaranteed. \n1028* exceeds the range in the above table, the correctness of the operator cannot be guaranteed. \n
1029-* In Ascend 950 AI Processor: If any dimension of x/filter/bias/offset_w/y shape exceeds max1029+* In Ascend 950PR&950DT products: If any dimension of x/filter/bias/offset_w/y shape exceeds max
1030* 1000000, the product of each dimension of x/filter/bias/offset_w/y1030* 1000000, the product of each dimension of x/filter/bias/offset_w/y
1031* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x1031* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x
1032* exceeds the range in the above table, the correctness of the operator cannot be guaranteed.1032* exceeds the range in the above table, the correctness of the operator cannot be guaranteed.
@@ -1068,8 +1068,8 @@ following types: float16.
1068 * @li x: A required 5D tensor of input image.1068 * @li x: A required 5D tensor of input image.
1069 The format of x is NCDHW or NDHWC.1069 The format of x is NCDHW or NDHWC.
1070 The data is stored in the order of: [n, in_channels, d, h, w] or [n, d, h, w, in_channels]. \n1070 The data is stored in the order of: [n, in_channels, d, h, w] or [n, d, h, w, in_channels]. \n
1071- Any dimension of x shape must be in [1, 2147483646] except Ascend 950 AI Processor. \n1071+ Any dimension of x shape must be in [1, 2147483646] except Ascend 950PR&950DT products. \n
1072- In Ascend 950 AI Processor, any dimension of x shape must be in [1, 1000000].1072+ In Ascend 950PR&950DT products, any dimension of x shape must be in [1, 1000000].
1073 * @li filter: A required 5D tensor of convolution kernel.1073 * @li filter: A required 5D tensor of convolution kernel.
1074 Must have the same type as "x".1074 Must have the same type as "x".
1075 The format support NCDHW or DHWCN.1075 The format support NCDHW or DHWCN.
@@ -1080,17 +1080,17 @@ following types: float16.
1080 The kernel_h and kernel_w dimensions must be in [1, 511]. \n1080 The kernel_h and kernel_w dimensions must be in [1, 511]. \n
1081 The other values of filter_size must be in [1, 2147483646]. \n1081 The other values of filter_size must be in [1, 2147483646]. \n
1082 When format is DHWCN and type is float32,1082 When format is DHWCN and type is float32,
1083- filter should be a constants except Ascend 950 AI Processor. \n1083+ filter should be a constants except Ascend 950PR&950DT products. \n
1084- In Ascend 950 AI Processor, the kernel_h and kernel_w dimensions must be in [1, 255],1084+ In Ascend 950PR&950DT products, the kernel_h and kernel_w dimensions must be in [1, 255],
1085 And the other values of filter_size must be in [1, 1000000].1085 And the other values of filter_size must be in [1, 1000000].
1086 * @li bias: An optional 1D tensor of additive biases to the outputs.1086 * @li bias: An optional 1D tensor of additive biases to the outputs.
1087 The data is stored in the order of: [out_channels].1087 The data is stored in the order of: [out_channels].
1088 "out_channels" must equals to the "out_channels" of output y. \n1088 "out_channels" must equals to the "out_channels" of output y. \n
1089- In Ascend 950 AI Processor, the out_channels dimension must be in [1, 1000000]1089+ In Ascend 950PR&950DT products, the out_channels dimension must be in [1, 1000000]
1090 * @li offset_w: An optional quantitative offset tensor. Reserved.1090 * @li offset_w: An optional quantitative offset tensor. Reserved.
1091*\n1091*\n
1092*\n1092*\n
1093-* The following are the supported data types and data formats for Ascend 950 AI Processor:1093+* The following are the supported data types and data formats for Ascend 950PR&950DT products:
1094*\n1094*\n
1095| Tensor | x | filter | bias | y |\n1095| Tensor | x | filter | bias | y |\n
1096| :-------: | :------: | :------: | :------: | :------: |\n1096| :-------: | :------: | :------: | :------: | :------: |\n
@@ -1116,19 +1116,19 @@ following types: float16.
1116 sliding window for each dimension of "x". The dimension order is determined by the data format of1116 sliding window for each dimension of "x". The dimension order is determined by the data format of
1117"x". The n and in_channels dimensions must be 1. \n When the format is "NDHWC", its shape is [1, stride_d, stride_h,1117"x". The n and in_channels dimensions must be 1. \n When the format is "NDHWC", its shape is [1, stride_d, stride_h,
1118stride_w, 1], when the format is "NCDHW", its shape is [1, 1, stride_d, stride_h, stride_w]. \n The stride_h and1118stride_w, 1], when the format is "NCDHW", its shape is [1, 1, stride_d, stride_h, stride_w]. \n The stride_h and
1119-stride_w dimensions must be in [1, 63]. The stride_d must be in [1, 2147483646] except Ascend 950 AI Processor.1119+stride_w dimensions must be in [1, 63]. The stride_d must be in [1, 2147483646] except Ascend 950PR&950DT products.
1120-\n In Ascend 950 AI Processor the stride_d must be in [1, 1000000].1120+\n In Ascend 950PR&950DT products the stride_d must be in [1, 1000000].
1121 * @li pads: Required. A list of 6 integers. Supports only padding along the d, h and w dimensions in sequence of1121 * @li pads: Required. A list of 6 integers. Supports only padding along the d, h and w dimensions in sequence of
1122 pad_head, pad_tail, pad_top, pad_bottom, pad_left and pad_right. \n1122 pad_head, pad_tail, pad_top, pad_bottom, pad_left and pad_right. \n
1123 The pad_top, pad_bottom, pad_left and pad_right must be in [0, 255].1123 The pad_top, pad_bottom, pad_left and pad_right must be in [0, 255].
1124- The pad_head and pad_tail must be in [0, 2147483646] except Ascend 950 AI Processor. \n1124+ The pad_head and pad_tail must be in [0, 2147483646] except Ascend 950PR&950DT products. \n
1125- In Ascend 950 AI Processor the pad_head and pad_tail must be in [1, 1000000].1125+ In Ascend 950PR&950DT products the pad_head and pad_tail must be in [1, 1000000].
1126 * @li dilations: Optional. A list of 5 integers. Specifies the dilation1126 * @li dilations: Optional. A list of 5 integers. Specifies the dilation
1127 factor for each dimension of "x". The dimension order is determined by the data format of "x".1127 factor for each dimension of "x". The dimension order is determined by the data format of "x".
1128\n When the format is "NDHWC", its shape is [1, dilation_d, dilation_h, dilation_w, 1], when the format is "NCDHW",1128\n When the format is "NDHWC", its shape is [1, dilation_d, dilation_h, dilation_w, 1], when the format is "NCDHW",
1129its shape is [1, 1, dilation_d, dilation_h, dilation_w]. \n Default value is [1, 1, 1, 1, 1]. \n The dilation_h and1129its shape is [1, 1, dilation_d, dilation_h, dilation_w]. \n Default value is [1, 1, 1, 1, 1]. \n The dilation_h and
1130-dilation_w dimensions must be in [1, 255]. The dilation_d dimensions must be in [0, 2147483646] except Ascend 950 AI1130+dilation_w dimensions must be in [1, 255]. The dilation_d dimensions must be in [0, 2147483646] except Ascend
1131-Processor. \n In Ascend 950 AI Processor the dilation_d dimensions must be in [1, 1000000].1131+950PR&950DT products. \n In Ascend 950PR&950DT products the dilation_d dimensions must be in [1, 1000000].
1132 * @li groups: Optional. An integer of type int32. The number of groups1132 * @li groups: Optional. An integer of type int32. The number of groups
1133 in group convolution. In_channels and out_channels must both be divisible by "groups".1133 in group convolution. In_channels and out_channels must both be divisible by "groups".
1134 The value of groups must be in [1, 65535]. Default value is 1.1134 The value of groups must be in [1, 65535]. Default value is 1.
@@ -1154,8 +1154,8 @@ Processor. \n In Ascend 950 AI Processor the dilation_d dimensions must be in [1
1154* (dilation_w * (kernel_w - 1) + 1))1154* (dilation_w * (kernel_w - 1) + 1))
1155* / stride_w + 11155* / stride_w + 1
1156*\n1156*\n
1157- Any dimension of y shape must be in [1, 2147483646] except Ascend 950 AI Processor. \n1157+ Any dimension of y shape must be in [1, 2147483646] except Ascend 950PR&950DT products. \n
1158- In Ascend 950 AI Processor, any dimension of y shape must be in [1, 1000000].1158+ In Ascend 950PR&950DT products, any dimension of y shape must be in [1, 1000000].
1159* @attention Constraints:1159* @attention Constraints:
1160 * @li The input x size after padding should be greater than the filter size.1160 * @li The input x size after padding should be greater than the filter size.
1161 * @li The w dimension of the input x supports cases exceeding 4096, but it may1161 * @li The w dimension of the input x supports cases exceeding 4096, but it may
@@ -1164,7 +1164,7 @@ Processor. \n In Ascend 950 AI Processor the dilation_d dimensions must be in [1
1164 * the product of each dimension of x/filter/bias/y shape exceeds max int32 minus one (2147483646) or1164 * the product of each dimension of x/filter/bias/y shape exceeds max int32 minus one (2147483646) or
1165 * the value of strides/pads/dilations/offset_x exceeds the range which is described in Attributes,1165 * the value of strides/pads/dilations/offset_x exceeds the range which is described in Attributes,
1166 * the correctness of the operator cannot be guaranteed. \n1166 * the correctness of the operator cannot be guaranteed. \n
1167- * In Ascend 950 AI Processor: If any dimension of x/filter/bias/y shape exceeds max1167+ * In Ascend 950PR&950DT products: If any dimension of x/filter/bias/y shape exceeds max
1168 * 1000000, the product of each dimension of x/filter/bias/y1168 * 1000000, the product of each dimension of x/filter/bias/y
1169 * shape exceeds max int32(2147483647) or the value of stride/padding/dilation/offset_x1169 * shape exceeds max int32(2147483647) or the value of stride/padding/dilation/offset_x
1170 * exceeds the range in the above table, the correctness of the operator cannot be guaranteed.1170 * exceeds the range in the above table, the correctness of the operator cannot be guaranteed.
@@ -1281,7 +1281,7 @@ Processor. \n In Ascend 950 AI Processor the dilation_d dimensions must be in [1
1281 | | W | [0, 4096] |\n1281 | | W | [0, 4096] |\n
1282 | Offset_x | | [-128, 127] |\n1282 | Offset_x | | [-128, 127] |\n
1283 *\n1283 *\n
1284- * In Atlas Training Series Product, fmap or out_backprop's H and W not support 1 when\n1284+ * In Atlas training products, fmap or out_backprop's H and W not support 1 when\n
1285 * fmap_h + pad_top + pad_bottom != (filter_height - 1) * dilation_h + 11285 * fmap_h + pad_top + pad_bottom != (filter_height - 1) * dilation_h + 1
1286 * and filter_width > fmap_width.1286 * and filter_width > fmap_width.
1287 * If filter_h = 1 and filter_w = 1, out_backprop_w * stride_h * stride_w1287 * If filter_h = 1 and filter_w = 1, out_backprop_w * stride_h * stride_w
@@ -1336,11 +1336,11 @@ Processor. \n In Ascend 950 AI Processor the dilation_d dimensions must be in [1
1336 * @li bias: Optional. An optional 1D tensor of type float16 and float32. When x1336 * @li bias: Optional. An optional 1D tensor of type float16 and float32. When x
1337 * is float16, bias is float16. When x is bfloat16, bias is float32.1337 * is float16, bias is float16. When x is bfloat16, bias is float32.
1338 * Currently bias is not supported on Atlas 200/500 A2 Inference Product and1338 * Currently bias is not supported on Atlas 200/500 A2 Inference Product and
1339- * Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component.1339+ * Atlas A2 products.
1340Reserved.1340Reserved.
1341 * @li offset_w: Optional. An optional 1D tensor for quantized deconvolution.1341 * @li offset_w: Optional. An optional 1D tensor for quantized deconvolution.
1342 * Currently offset_w is not supported on Atlas 200/500 A2 Inference Product and1342 * Currently offset_w is not supported on Atlas 200/500 A2 Inference Product and
1343- * Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component.1343+ * Atlas A2 products.
1344Reserved.1344Reserved.
1345 1345 
1346*@par Attributes:1346*@par Attributes:
@@ -1367,11 +1367,11 @@ Reserved.
1367 * @li output_padding: Optional. The size will be added in the output shape.1367 * @li output_padding: Optional. The size will be added in the output shape.
1368 * Defaults to [0, 0, 0, 0, 0].1368 * Defaults to [0, 0, 0, 0, 0].
1369 * Currently output_padding is not supported on Atlas 200/500 A2 Inference Product and1369 * Currently output_padding is not supported on Atlas 200/500 A2 Inference Product and
1370- * Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component.1370+ * Atlas A2 products.
1371 * In graph mode, only configuration to [0, 0, 0, 0, 0] is supported.1371 * In graph mode, only configuration to [0, 0, 0, 0, 0] is supported.
1372 * @li offset_x: Optional. Input offset_x value. Defaults to 0.1372 * @li offset_x: Optional. Input offset_x value. Defaults to 0.
1373 * Currently offset_x is not supported on Atlas 200/500 A2 Inference Product and1373 * Currently offset_x is not supported on Atlas 200/500 A2 Inference Product and
1374- * Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component.1374+ * Atlas A2 products.
1375Reserved.1375Reserved.
1376 1376 
1377*@par Outputs:1377*@par Outputs:
@@ -1431,12 +1431,13 @@ Reserved.
1431 * @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.1431 * @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.
1432 * The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.1432 * The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.
1433 * For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n1433 * For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n
1434- * For Non-Ascend 950 AI Processor: The produce of the ksize in H and W dimensions1434+ * For products other than Ascend 950PR&950DT products: The produce of the ksize in H and W dimensions
1435 * should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n1435 * should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n
1436 * @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n1436 * @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n
1437- * For Non-Ascend 950 AI Processor: The stride of the H and W dimensions should be greater than 0 and1437+ * For products other than Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0
1438+ and
1438 * smaller than 64. \n1439 * smaller than 64. \n
1439- * For Ascend 950 AI Processor: The stride of the H and W dimensions should be greater than 0.1440+ * For Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0.
1440 * @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The1441 * @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The
1441 output1442 output
1442 * shape at the H and W dimensions is calculated by below formula: \n1443 * shape at the H and W dimensions is calculated by below formula: \n
@@ -2316,7 +2317,7 @@ currently supported.
2316 * Format of the original input, either NCHW or NHWC. Defaults to NHWC . \n2317 * Format of the original input, either NCHW or NHWC. Defaults to NHWC . \n
2317 2318 
2318 * @attention Constraints:2319 * @attention Constraints:
2319- * @li Only Atlas Training Series Product is supported.2320+ * @li Only Atlas training products are supported.
2320 * @li "x1" and "grads" must have the same shape.2321 * @li "x1" and "grads" must have the same shape.
2321 * @li "x2" and "y" must have the same shape. Otherwise, an error is reported.2322 * @li "x2" and "y" must have the same shape. Otherwise, an error is reported.
2322 * @li "x1", "x2", "grads", and "y" must be 5D tensors.2323 * @li "x1", "x2", "grads", and "y" must be 5D tensors.
@@ -2434,7 +2435,7 @@ currently supported.
2434 * Must set the format, supported format list ["NC1HWC0"]2435 * Must set the format, supported format list ["NC1HWC0"]
2435 * @li argmax: An 5hd tensor of type uint16 or int64.2436 * @li argmax: An 5hd tensor of type uint16 or int64.
2436 * Must set the format, supported format list ["NC1HWC0"] \n2437 * Must set the format, supported format list ["NC1HWC0"] \n
2437- * For Ascend 950 AI Processor: The uint16 data type is not supported.2438+ * For Ascend 950PR&950DT products: The uint16 data type is not supported.
2438 2439 
2439 * @par Attributes:2440 * @par Attributes:
2440 * @li ksize: A required list of int8, int16, int32, or int64 values,2441 * @li ksize: A required list of int8, int16, int32, or int64 values,
@@ -2778,7 +2779,7 @@ currently supported.
2778 2779 
2779 *@attention Constraints:2780 *@attention Constraints:
2780 *@li UniqueWithCounts runs on the Ascend AI CPU, which delivers poor performance. \n2781 *@li UniqueWithCounts runs on the Ascend AI CPU, which delivers poor performance. \n
2781- *@li Dtype bfloat16, uint32, uint64 only support Ascend 950 AI Processor. \n2782+ *@li Dtype bfloat16, uint32, uint64 only support Ascend 950PR&950DT products. \n
2782 2783 
2783 *@par Third-party framework compatibility2784 *@par Third-party framework compatibility
2784 *Compatible with the TensorFlow operator UniqueWithCounts.2785 *Compatible with the TensorFlow operator UniqueWithCounts.
@@ -2809,7 +2810,7 @@ currently supported.
2809 2810 
2810 *@attention Constraints:2811 *@attention Constraints:
2811 *@li Unique runs on the Ascend AI CPU, which delivers poor performance. \n2812 *@li Unique runs on the Ascend AI CPU, which delivers poor performance. \n
2812- *@li Dtype bfloat16, uint32, uint64 only support Ascend 950 AI Processor. \n2813+ *@li Dtype bfloat16, uint32, uint64 only support Ascend 950PR&950DT products. \n
2813 2814 
2814 *@par Third-party framework compatibility2815 *@par Third-party framework compatibility
2815 *Compatible with the TensorFlow operator Unique.2816 *Compatible with the TensorFlow operator Unique.
@@ -2917,8 +2918,6 @@ currently supported.
2917 * @li This operator is a BatchNorm fusion operator for updating the moving2918 * @li This operator is a BatchNorm fusion operator for updating the moving
2918 * averages for training. This operator is used in conjunction with2919 * averages for training. This operator is used in conjunction with
2919 * BNTrainingUpdate.2920 * BNTrainingUpdate.
2920- * @li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the
2921- * square root instruction.
2922 */2921 */
2923 REG_OP(BNTrainingUpdate)2922 REG_OP(BNTrainingUpdate)
2924 .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))2923 .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
@@ -3099,8 +3098,6 @@ currently supported.
3099 *@attention Constraints:3098 *@attention Constraints:
3100 *@li This operator is a InstanceNorm fusion operator for updating the moving averages for training.3099 *@li This operator is a InstanceNorm fusion operator for updating the moving averages for training.
3101 * This operator is used in conjunction with GNTrainingUpdate.3100 * This operator is used in conjunction with GNTrainingUpdate.
3102- *@li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the square root
3103- instruction.
3104 */3101 */
3105 REG_OP(GNTrainingUpdate)3102 REG_OP(GNTrainingUpdate)
3106 .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))3103 .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
Mcommon/stub/op_graph/math_proto_stub.cpp+3-3
@@ -60,13 +60,13 @@ REG_OP(Add)
60 * If the input data contains -inf, -inf is converted into the minimum value of int16. \n60 * If the input data contains -inf, -inf is converted into the minimum value of int16. \n
61 * @li In the scenarios where the data type is converted from INT32 to INT8: \n61 * @li In the scenarios where the data type is converted from INT32 to INT8: \n
62 * It can only guarantee that the input data has no precision errors within the range of (-2048, 1920).62 * It can only guarantee that the input data has no precision errors within the range of (-2048, 1920).
63- * @li Atlas Inference Series Product in the scenarios where the data type is converted from FLOAT32 to INT8: \n63+ * @li Atlas inference products in the scenarios where the data type is converted from FLOAT32 to INT8: \n
64 * It can only guarantee that the input data has no precision errors within the range of (-2048, 1920).64 * It can only guarantee that the input data has no precision errors within the range of (-2048, 1920).
65- * @li Atlas Inference Series Product in the scenarios where the data type is converted from FLOAT32 to INT64 and65+ * @li Atlas inference products in the scenarios where the data type is converted from FLOAT32 to INT64 and
66 from FLOAT32 to UINT8: \n66 from FLOAT32 to UINT8: \n
67 * It can only guarantee that the input data has no precision errors within the range of (-2147483648,67 * It can only guarantee that the input data has no precision errors within the range of (-2147483648,
68 2147483583).68 2147483583).
69- * @li Atlas Inference Series Product in the scenarios where the data type is converted from INT64 to FLOAT32: \n69+ * @li Atlas inference products in the scenarios where the data type is converted from INT64 to FLOAT32: \n
70 * It can only guarantee that the input data has no precision errors within the range of (-2147483648,70 * It can only guarantee that the input data has no precision errors within the range of (-2147483648,
71 2147483647).71 2147483647).
72 */72 */
Mconv/conv3d_backprop_filter_v2/op_graph/conv3d_backprop_filter_v2_proto.h+3-2
@@ -73,7 +73,8 @@ namespace ge {
73*@par Outputs:73*@par Outputs:
74 * y: A Tensor that has the type float32. The format is NCDHW or NDHWC or DHWCN.74 * y: A Tensor that has the type float32. The format is NCDHW or NDHWC or DHWCN.
75* @attention Constraints:75* @attention Constraints:
76- * In Ascend 950PR/Ascend 950DT: The behavior of gradient computation in the padding region depends on the input shape.76+ * In Ascend 950PR&950DT products: The behavior of gradient computation in the padding region depends on the input
77+shape.
77 * Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n78 * Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n
78*@par Third-party framework compatibility79*@par Third-party framework compatibility
79 * Compatible with Tensorflow's conv3d_backprop_filter80 * Compatible with Tensorflow's conv3d_backprop_filter
@@ -92,4 +93,4 @@ REG_OP(Conv3DBackpropFilterV2)
92 .OP_END_FACTORY_REG(Conv3DBackpropFilterV2)93 .OP_END_FACTORY_REG(Conv3DBackpropFilterV2)
93} // namespace ge94} // namespace ge
94 95 
95-#endif // OPS_CONV_CONV3D_BACKPROP_FILTER_PROTO_H_96+#endif // OPS_CONV_CONV3D_BACKPROP_FILTER_PROTO_H_
Mconv/conv3d_backprop_input_v2/op_graph/conv3d_backprop_input_v2_proto.h+2-2
@@ -89,8 +89,8 @@ namespace ge {
89 * The type is float16, bfloat16, float32.89 * The type is float16, bfloat16, float32.
90 * The gradients of feature map.90 * The gradients of feature map.
91 * @attention Constraints:91 * @attention Constraints:
92- * In Ascend 950PR/Ascend 950DT: The behavior of gradient computation in the padding region depends on the input shape.92+ * In Ascend 950PR&950DT products: The behavior of gradient computation in the padding region depends on the input
93- * Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n93+ *shape. Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n
94 *@par Third-party framework compatibility94 *@par Third-party framework compatibility
95 * Compatible with Tensorflow's conv3d_backprop_input95 * Compatible with Tensorflow's conv3d_backprop_input
96 */96 */
Mconv/conv3d_transpose_v2/op_graph/conv3d_transpose_v2_proto.h+2-1
@@ -90,7 +90,8 @@ namespace ge {
90*@par Outputs:90*@par Outputs:
91 * y: A tensor that has the type bfloat16, float16, float32, hifloat8, float8_e4m3fn. It has the same format as "x".91 * y: A tensor that has the type bfloat16, float16, float32, hifloat8, float8_e4m3fn. It has the same format as "x".
92 * @attention Constraints:92 * @attention Constraints:
93- * In Ascend 950PR/Ascend 950DT: The behavior of gradient computation in the padding region depends on the input shape.93+ * In Ascend 950PR&950DT products: The behavior of gradient computation in the padding region depends on the input
94+shape.
94 * Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n95 * Depending on the operator optimization strategy, the padding gradients may be directly set to 0.\n
95*/96*/
96REG_OP(Conv3DTransposeV2)97REG_OP(Conv3DTransposeV2)
Mconv/conv3d_v2/op_graph/conv3d_v2_proto.h+27-27
@@ -22,46 +22,46 @@ namespace ge {
22/**22/**
23* @brief Computes a 3D convolution with 5D "x", "filter" and "bias" tensors.23* @brief Computes a 3D convolution with 5D "x", "filter" and "bias" tensors.
24* Like this, output = CONV(x, filter) + bias. \n24* Like this, output = CONV(x, filter) + bias. \n
25-* If case with 'int8' dtype appears in Atlas A3 Training Series Product/Atlas A3 Inference Series Product or25+* If case with 'int8' dtype appears in Atlas A3 products or
26-* Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component,26+* Atlas A2 products,
27* like this: output = CONV(x, filter) * scale + bias.27* like this: output = CONV(x, filter) * scale + bias.
28* @par Inputs:28* @par Inputs:
29* @li x: A required 5D tensor of input image. \n29* @li x: A required 5D tensor of input image. \n
30-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or30+* In Atlas A2 products or
31-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,31+* Atlas A3 products,
32* a tensor with data type bfloat16, float16, float32, int8 and format "NCDHW" is supported. \n32* a tensor with data type bfloat16, float16, float32, int8 and format "NCDHW" is supported. \n
33-* In Ascend 950 AI Processor, a tensor of type bfloat16, float16, float32 or hifloat8 and format "NCDHW" or33+* In Ascend 950PR&950DT products, a tensor of type bfloat16, float16, float32 or hifloat8 and format "NCDHW" or
34* "NDHWC" can be supported.34* "NDHWC" can be supported.
35* @li filter: A required 5D tensor of convolution kernel. \n35* @li filter: A required 5D tensor of convolution kernel. \n
36-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or36+* In Atlas A2 products or
37-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,37+* Atlas A3 products,
38* a tensor with data type bfloat16, float16, float32, int8 and format "NCDHW" is supported.38* a tensor with data type bfloat16, float16, float32, int8 and format "NCDHW" is supported.
39* Kernel_h and kernel_w should be both less than 512. \n39* Kernel_h and kernel_w should be both less than 512. \n
40-* In Ascend 950 AI Processor, a tensor with data type bfloat16, float16, float32 or hifloat8 and format "NCDHW" or40+* In Ascend 950PR&950DT products, a tensor with data type bfloat16, float16, float32 or hifloat8 and format "NCDHW" or
41* "DHWCN" can be supported. Kernel_h and kernel_w should be both less than 256.41* "DHWCN" can be supported. Kernel_h and kernel_w should be both less than 256.
42* @li bias: An optional 1D tensor of additive biases to the outputs.42* @li bias: An optional 1D tensor of additive biases to the outputs.
43* The data is stored in the order of: [out_channels]. \n43* The data is stored in the order of: [out_channels]. \n
44-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or44+* In Atlas A2 products or
45-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,45+* Atlas A3 products,
46* a tensor with data type float16, float32 and format "ND" is supported. \n46* a tensor with data type float16, float32 and format "ND" is supported. \n
47-* In Ascend 950 AI Processor, a tensor with data type bfloat16, float16 or float32 and format "ND" can be supported.47+* In Ascend 950PR&950DT products, a tensor with data type bfloat16, float16 or float32 and format "ND" can be supported.
48* @li scale: A optional 1D tensor of scaling factors.48* @li scale: A optional 1D tensor of scaling factors.
49* The data is stored in the order of: [out_channels]. \n49* The data is stored in the order of: [out_channels]. \n
50-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or50+* In Atlas A2 products or
51-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,51+* Atlas A3 products,
52* a tensor with data type float32 and format "ND" is supported. \n52* a tensor with data type float32 and format "ND" is supported. \n
53-* In Ascend 950 AI Processor, this parameter is not supported.53+* In Ascend 950PR&950DT products, this parameter is not supported.
54* @li offset: An optional 1D tensor of bias.54* @li offset: An optional 1D tensor of bias.
55* The data is stored in the order of: [out_channels].55* The data is stored in the order of: [out_channels].
56-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or56+* In Atlas A2 products or
57-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,57+* Atlas A3 products,
58* a tensor with data type float32 and format "ND" is supported. \n58* a tensor with data type float32 and format "ND" is supported. \n
59-* In Ascend 950 AI Processor, this parameter is not supported.59+* In Ascend 950PR&950DT products, this parameter is not supported.
60* @li offset_w: An optional quantitative offset tensor. A tensor of type int8. Reserved.60* @li offset_w: An optional quantitative offset tensor. A tensor of type int8. Reserved.
61*\n61*\n
62* @li The following are the supported data types and data formats for62* @li The following are the supported data types and data formats for
63-* (Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component and63+* (Atlas A2 products and
64-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product):64+* Atlas A3 products):
65\n65\n
66| Tensor | x | filter | bias | scale | offset | y |\n66| Tensor | x | filter | bias | scale | offset | y |\n
67| :-------: | :------: | :------: | :------: | :-----: | :-----: | :------: |\n67| :-------: | :------: | :------: | :------: | :-----: | :-----: | :------: |\n
@@ -71,7 +71,7 @@ namespace ge {
71| | int8 | int8 | float32 | float32 | float32 | bfloat16 |\n71| | int8 | int8 | float32 | float32 | float32 | bfloat16 |\n
72| Format | NCDHW | NCDHW | ND | ND | ND | NCDHW |\n72| Format | NCDHW | NCDHW | ND | ND | ND | NCDHW |\n
73\n73\n
74-* The following are the supported data types and data formats for Ascend 950 AI Processor:74+* The following are the supported data types and data formats for Ascend 950PR&950DT products:
75\n75\n
76| Tensor | x | filter | bias | y |\n76| Tensor | x | filter | bias | y |\n
77| :-------: | :------: | :------: | :------: | :------: |\n77| :-------: | :------: | :------: | :------: | :------: |\n
@@ -97,24 +97,24 @@ namespace ge {
97* when the format is "NCDHW", its shape is [1, 1, dilation_d, dilation_h, dilation_w]. Defaults to [1, 1, 1, 1].97* when the format is "NCDHW", its shape is [1, 1, dilation_d, dilation_h, dilation_w]. Defaults to [1, 1, 1, 1].
98* @li groups: Optional. An integer of type int32. The number of groups98* @li groups: Optional. An integer of type int32. The number of groups
99* in group convolution. In_channels and out_channels must both be divisible by "groups". Defaults to 1.99* in group convolution. In_channels and out_channels must both be divisible by "groups". Defaults to 1.
100-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or100+* In Atlas A2 products or
101-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product, groups can only be equal to 1.101+* Atlas A3 products, groups can only be equal to 1.
102* @li data_format: Optional. It is a string represents input's data format. Defaults to "NCDHW". Reserved.102* @li data_format: Optional. It is a string represents input's data format. Defaults to "NCDHW". Reserved.
103* @li offset_x: Optional. An integer of type int32. It means offset in quantization algorithm103* @li offset_x: Optional. An integer of type int32. It means offset in quantization algorithm
104* and is used for filling in pad values. Defaults to 0. It can only be supported in104* and is used for filling in pad values. Defaults to 0. It can only be supported in
105-* Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or105+* Atlas A2 products or
106-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product.106+* Atlas A3 products.
107* @li pad_mode: Optional. An optional string parameter, indicating the mode of pad.107* @li pad_mode: Optional. An optional string parameter, indicating the mode of pad.
108* It must be "SPECIFIC" or "SAME" or "VALID". Defaults to "SPECIFIC".108* It must be "SPECIFIC" or "SAME" or "VALID". Defaults to "SPECIFIC".
109* @li enable_hf32: Optional. An optional bool parameter. Used to enable hf32 computation.109* @li enable_hf32: Optional. An optional bool parameter. Used to enable hf32 computation.
110* If true, enable hf32 computation, otherwise, disable hf32 computation. Defaults to false.110* If true, enable hf32 computation, otherwise, disable hf32 computation. Defaults to false.
111* @par Outputs:111* @par Outputs:
112* y: A 5D tensor of output. \n112* y: A 5D tensor of output. \n
113-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or113+* In Atlas A2 products or
114-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,114+* Atlas A3 products,
115* a tensor with data type bfloat16,float32,float16 and format "NCDHW" is supported,115* a tensor with data type bfloat16,float32,float16 and format "NCDHW" is supported,
116* which the data is stored in [n, out_channels, out_depth, out_height, out_width]. \n116* which the data is stored in [n, out_channels, out_depth, out_height, out_width]. \n
117-* In Ascend 950 AI Processor, a tensor with data type bfloat16, float16, float32 or hifloat8 and117+* In Ascend 950PR&950DT products, a tensor with data type bfloat16, float16, float32 or hifloat8 and
118* format "NCDHW" or "NDHWC" can be supported. which the data is stored in118* format "NCDHW" or "NDHWC" can be supported. which the data is stored in
119* [n, out_channels, out_depth, out_height, out_width] or [n, out_depth, out_height, out_width, out_channels].119* [n, out_channels, out_depth, out_height, out_width] or [n, out_depth, out_height, out_width, out_channels].
120*\n120*\n
Mconv/extend_conv2d/op_graph/extend_conv2d_proto.h+1-1
@@ -149,7 +149,7 @@ float8 | float32/float16/bfloat16/float8_e4m3 |\n | Format | NCHW
149| enable_relu1 | - | [true, false] |\n149| enable_relu1 | - | [true, false] |\n
150| dual_output | - | [true, false] |\n150| dual_output | - | [true, false] |\n
151*\n151*\n
152-* @li In Ascend 950 AI Processor: If any dimension of152+* @li In Ascend 950PR&950DT products: If any dimension of
153* x/filter/bias/offset_w/scale0/relu_weight0/clip_value0/scale1/relu_weight1/clip_value1/y shape exceeds max153* x/filter/bias/offset_w/scale0/relu_weight0/clip_value0/scale1/relu_weight1/clip_value1/y shape exceeds max
154* 1000000, the product of each dimension of x/filter/bias/scale/offset/y154* 1000000, the product of each dimension of x/filter/bias/scale/offset/y
155* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x155* shape exceeds max int32(2147483647) or the value of strides/pads/dilations/offset_x
Mconv/quant_conv3d/op_graph/quant_conv3d_proto.h+1-1
@@ -34,7 +34,7 @@ namespace ge {
34* The data is stored in the order of: [out_channels].34* The data is stored in the order of: [out_channels].
35* @li offset: An optional quantitative offset tensor. Reserved.35* @li offset: An optional quantitative offset tensor. Reserved.
36*\n36*\n
37-* The following are the supported data types and data formats (for Ascend 950 AI Processor):37+* The following are the supported data types and data formats (for Ascend 950PR&950DT products):
38*\n38*\n
39| Tensor | x | filter | scale | bias | offset | y |\n39| Tensor | x | filter | scale | bias | offset | y |\n
40| :-------: | :---------: | :---------: | :----------: | :------:| :------:| :---------------------------------: |\n40| :-------: | :---------: | :---------: | :----------: | :------:| :------:| :---------------------------------: |\n
Mexperimental/index/index_fill/op_graph/index_fill_proto.h+2-6
@@ -22,17 +22,13 @@ namespace ge {
22* @par Inputs:22* @par Inputs:
23* Three inputs, including:23* Three inputs, including:
24* @li x: A tensor that serves as the source; its duplicate is created and subsequently filled with the specified values.24* @li x: A tensor that serves as the source; its duplicate is created and subsequently filled with the specified values.
25-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or Atlas A325+* In Atlas A2 products or Atlas A3 products, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8,
26-Training Series
27-* Product/Atlas A3 Inference Series Product, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8,
28uint8, int16, double can be supported. \n26uint8, int16, double can be supported. \n
29* @li indices: A tensor, which equivalent to a vector or scalar. indices of input tensor to fill in. Must be one of the27* @li indices: A tensor, which equivalent to a vector or scalar. indices of input tensor to fill in. Must be one of the
30following types:28following types:
31* int32, int64. \n29* int32, int64. \n
32* @li val: The value to fill with. It's a scalar or a one-dimensional tensor with only one element.30* @li val: The value to fill with. It's a scalar or a one-dimensional tensor with only one element.
33-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or Atlas A331+* In Atlas A2 products or Atlas A3 products, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8,
34-Training Series
35-* Product/Atlas A3 Inference Series Product, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8,
36uint8, int16, double can be supported. \n32uint8, int16, double can be supported. \n
37* @par Attributes:33* @par Attributes:
38* dim: A required int. Used to select the dimension of the input tensor. \n34* dim: A required int. Used to select the dimension of the input tensor. \n
Mindex/embedding_dense_grad_v2/op_graph/embedding_dense_grad_v2_proto.h+1-1
@@ -45,7 +45,7 @@ namespace ge {
45 45 
46* @par Restrictions:46* @par Restrictions:
47* Warning: The grad of input last dim size should be less than 8192 bytes and the47* Warning: The grad of input last dim size should be less than 8192 bytes and the
48-* sort_indices size should be larger than 32 for Ascend 950 AI Processors. \n48+* sort_indices size should be larger than 32 for Ascend 950PR&950DT products. \n
49*/49*/
50REG_OP(EmbeddingDenseGradV2)50REG_OP(EmbeddingDenseGradV2)
51 .INPUT(grad, TensorType({DT_FLOAT32, DT_FLOAT16, DT_BF16})) /* "First operand." */51 .INPUT(grad, TensorType({DT_FLOAT32, DT_FLOAT16, DT_BF16})) /* "First operand." */
Mindex/index_fill/op_graph/index_fill_proto.h+5-9
@@ -22,21 +22,17 @@ namespace ge {
22* @par Inputs:22* @par Inputs:
23* Three inputs, including:23* Three inputs, including:
24* @li x: A tensor that serves as the source; its duplicate is created and subsequently filled with the specified values.24* @li x: A tensor that serves as the source; its duplicate is created and subsequently filled with the specified values.
25-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or Atlas A325+* In Atlas A2 products or Atlas A3 products, a tensor of type float16, float32, bfloat16, int64, int32, bool can be
26-Training Series
27-* Product/Atlas A3 Inference Series Product, a tensor of type float16, float32, bfloat16, int64, int32, bool can be
28supported. \n26supported. \n
29-* In Ascend 950 AI Processor, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8, uint8, int16,27+* In Ascend 950PR&950DT products, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8, uint8, int16,
30double can be supported. \n28double can be supported. \n
31* @li indices: A tensor, which equivalent to a vector or scalar. indices of input tensor to fill in. Must be one of the29* @li indices: A tensor, which equivalent to a vector or scalar. indices of input tensor to fill in. Must be one of the
32following types:30following types:
33* int32, int64. \n31* int32, int64. \n
34* @li val: The value to fill with. It's a scalar or a one-dimensional tensor with only one element.32* @li val: The value to fill with. It's a scalar or a one-dimensional tensor with only one element.
35-* In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or Atlas A333+* In Atlas A2 products or Atlas A3 products, a tensor of type float16, float32, bfloat16, int64, int32, bool can be
36-Training Series
37-* Product/Atlas A3 Inference Series Product, a tensor of type float16, float32, bfloat16, int64, int32, bool can be
38supported. \n34supported. \n
39-* In Ascend 950 AI Processor, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8, uint8, int16,35+* In Ascend 950PR&950DT products, a tensor of type float16, float32, bfloat16, int64, int32, bool, int8, uint8, int16,
40double can be supported. \n36double can be supported. \n
41* @par Attributes:37* @par Attributes:
42* dim: A required int. Used to select the dimension of the input tensor. \n38* dim: A required int. Used to select the dimension of the input tensor. \n
@@ -66,4 +62,4 @@ REG_OP(IndexFill)
66 .OP_END_FACTORY_REG(IndexFill)62 .OP_END_FACTORY_REG(IndexFill)
67} // namespace ge63} // namespace ge
68 64 
69-#endif // OPS_BUILT_IN_OP_PROTO_INC_INDEX_FILL_H_65+#endif // OPS_BUILT_IN_OP_PROTO_INC_INDEX_FILL_H_
Mindex/map_index/op_graph/map_index_proto.h+4-4
@@ -27,11 +27,11 @@ namespace ge {
27* @par Inputs:27* @par Inputs:
28* Three inputs, including:28* Three inputs, including:
29* @li x: One dimensional tensor of type int32, specifying queried shape,29* @li x: One dimensional tensor of type int32, specifying queried shape,
30-* Format support ND, the max dim is 400 (128 on Lhisi(Hi3796CV300CS, SD3403), 24000 on Ascend 910_95 AI Processor).30+* Format support ND, the max dim is 400 (128 on Lhisi(Hi3796CV300CS, SD3403), 24000 on Ascend 950PR&950DT products).
31* @li data_seq: One dimensional tensor of type int32,31* @li data_seq: One dimensional tensor of type int32,
32* Format support ND, specifying the mapped table is queried.32* Format support ND, specifying the mapped table is queried.
33-* The length of data_seq must be multiple of the length of x, and the length of data_seq / x <= 100 (256 on Ascend33+* The length of data_seq must be multiple of the length of x, and the length of data_seq / x <= 100 (256 on
34-910_95 AI Processor).34+* Ascend 950PR&950DT products).
35* @li level_index: One dimensional tensor of type int32, the length of level_index must be equal to the length of35* @li level_index: One dimensional tensor of type int32, the length of level_index must be equal to the length of
36data_seq divided by the length of x.36data_seq divided by the length of x.
37* Format support ND, specifying secondary index. \n37* Format support ND, specifying secondary index. \n
@@ -54,4 +54,4 @@ REG_OP(MapIndex)
54 54 
55} // namespace ge55} // namespace ge
56 56 
57-#endif // OPS_OP_PROTO_INC_MAP_INDEX_H_57+#endif // OPS_OP_PROTO_INC_MAP_INDEX_H_
Mindex/non_zero_with_value/op_graph/non_zero_with_value_proto.h+2-2
@@ -31,10 +31,10 @@ namespace ge {
31* uint32, int64, uint64, bool. \n31* uint32, int64, uint64, bool. \n
32 32 
33* @par Attributes:33* @par Attributes:
34-* @li transpose: An optional bool. Defaults to false. On <term>Ascend 950</term> (arch35) only34+* @li transpose: An optional bool. Defaults to false. On Ascend 950PR&950DT products (arch35) only
35* transpose=true (coordinate-major index layout) is supported; false is not delivered.35* transpose=true (coordinate-major index layout) is supported; false is not delivered.
36* @li dtype: An optional attribute specifying the output index data type. Defaults to DT_INT32.36* @li dtype: An optional attribute specifying the output index data type. Defaults to DT_INT32.
37-* On <term>Ascend 950</term> (arch35) only int32 is supported. \n37+* On Ascend 950PR&950DT products (arch35) only int32 is supported. \n
38 38 
39* @par Outputs:39* @par Outputs:
40* @li value: A Tensor. Has the same type as "x". Shape is [row*col] (static max-size);40* @li value: A Tensor. Has the same type as "x". Shape is [row*col] (static max-size);
Mindex/quant_update_scatter/op_graph/quant_update_scatter_proto.h+2-3
@@ -58,8 +58,7 @@ to "rint".
58* @attention Constraints:58* @attention Constraints:
59* @li updates，quant_scales，quant_zero_points dtype should be one of the following59* @li updates，quant_scales，quant_zero_points dtype should be one of the following
60combinations:[(BFLOAT16，BFLOAT16，BFLOAT16), (FLOAT16，FLOAT32，INT32)]60combinations:[(BFLOAT16，BFLOAT16，BFLOAT16), (FLOAT16，FLOAT32，INT32)]
61-* @li Atlas Inference Series Product and Atlas Trainning Series Product and Atlas A2 Training Series Product/Atlas 800I61+* @li Atlas inference products and Atlas training products and Atlas A2 products and Atlas A3 training products:
62-A2 Inference Product and Atlas A3 Training Series Product:
63* Output param var last dim should be 32B-aligned.62* Output param var last dim should be 32B-aligned.
64* @li When var is DT_INT8, DT_FLOAT8_E5M2 or DT_FLOAT8_E4M3FN, round_mode only supports "rint".63* @li When var is DT_INT8, DT_FLOAT8_E5M2 or DT_FLOAT8_E4M3FN, round_mode only supports "rint".
65* @li When var is DT_HIFLOAT8, round_mode supports "round" and "hybrid".64* @li When var is DT_HIFLOAT8, round_mode supports "round" and "hybrid".
@@ -81,4 +80,4 @@ REG_OP(QuantUpdateScatter)
81 80 
82} // namespace ge81} // namespace ge
83 82 
84-#endif // OPS_OP_PROTO_INC_MASKED_SCATTER_H_83+#endif // OPS_OP_PROTO_INC_MASKED_SCATTER_H_
Mindex/scatter_elements/op_graph/scatter_elements_proto.h+2-2
@@ -37,8 +37,8 @@ qint32, qint8, quint8, uint16, uint32, uint64, uint8, bfloat16, complex32.
37* y: A Tensor. Has the same type and format as input "data" . \n37* y: A Tensor. Has the same type and format as input "data" . \n
38 38 
39* @attention Constraints:39* @attention Constraints:
40-* @li In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component and40+* @li In Atlas A2 products and
41-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,41+* Atlas A3 products,
42* you are advised to replace ScatterElements with ScatterElementsV2(When there are duplicate indexes, ScatterElementsV242* you are advised to replace ScatterElements with ScatterElementsV2(When there are duplicate indexes, ScatterElementsV2
43provides higher precision). \n43provides higher precision). \n
44 44 
Mindex/scatter_elements_v2/op_graph/scatter_elements_v2_proto.h+2-2
@@ -41,8 +41,8 @@ namespace ge {
41* @attention Constraints:41* @attention Constraints:
42* @li In non-last axis scenarios, you are advised to convert x, indices, and updates to the last axes,42* @li In non-last axis scenarios, you are advised to convert x, indices, and updates to the last axes,
43* use ScatterElementsV2 for calculation, and then convert them to the original axes.43* use ScatterElementsV2 for calculation, and then convert them to the original axes.
44-* @li Only Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component and44+* @li Only Atlas A2 products and
45-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product support ScatterElementsV2. \n45+* Atlas A3 products support ScatterElementsV2. \n
46 46 
47* @par Outputs:47* @par Outputs:
48* var: A Tensor. Has the same type and format as input "var" . \n48* var: A Tensor. Has the same type and format as input "var" . \n
Mindex/sparse_to_dense/op_graph/sparse_to_dense_proto.h+1-1
@@ -41,7 +41,7 @@ double.
41* @li validate_indices: An optional bool.41* @li validate_indices: An optional bool.
42* If true, indices are checked to make sure they are sorted by ascending, no repeats, and cannot exceed the size of each42* If true, indices are checked to make sure they are sorted by ascending, no repeats, and cannot exceed the size of each
43dimension.43dimension.
44-* This param is currently not effective in Ascend950PR/Ascend950DT. \n44+* This param is currently not effective in Ascend 950PR&950DT products. \n
45 45 
46* @par Outputs:46* @par Outputs:
47* y: A Tensor. Has the same type and format as input "values" . \n47* y: A Tensor. Has the same type and format as input "values" . \n
Mloss/l1_loss_grad/op_graph/l1_loss_grad_proto.h+3-3
@@ -34,8 +34,8 @@ namespace ge {
34* y: A Tensor. Has the same type as "grads". \n34* y: A Tensor. Has the same type as "grads". \n
35 35 
36* @attention Constraints:36* @attention Constraints:
37-* @li In Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component and37+* @li In Atlas A2 products and
38-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product,38+* Atlas A3 products,
39* broadcasting is not allowed between input grads, input predict and input label, the three inputs must have the same39* broadcasting is not allowed between input grads, input predict and input label, the three inputs must have the same
40shape. \n40shape. \n
41 41 
@@ -50,4 +50,4 @@ REG_OP(L1LossGrad)
50 .ATTR(reduction, String, "mean")50 .ATTR(reduction, String, "mean")
51 .OP_END_FACTORY_REG(L1LossGrad)51 .OP_END_FACTORY_REG(L1LossGrad)
52} // namespace ge52} // namespace ge
53-#endif // L1_LOSS_GRAD_OP_PROTO_H_53+#endif // L1_LOSS_GRAD_OP_PROTO_H_
Mloss/softmax_focal_loss/op_graph/softmax_focal_loss_proto.h+1-1
@@ -26,7 +26,7 @@ namespace ge {
26*Three inputs, including:26*Three inputs, including:
27* @li pred: A Tensor. Must be one of the following types: float16, float32.27* @li pred: A Tensor. Must be one of the following types: float16, float32.
28*The probabilities produced by a preceding softmax. The last axis is the class axis and the28*The probabilities produced by a preceding softmax. The last axis is the class axis and the
29-*remaining axes are sample axes. On <term>Ascend 950PR&950DT系列产品</term> any rank >= 1 is29+*remaining axes are sample axes. On Ascend 950PR&950DT products any rank >= 1 is
30*accepted; the other products only support the two-dimensional form "batch_size * num_classes".30*accepted; the other products only support the two-dimensional form "batch_size * num_classes".
31* @li target: A Tensor of type int32. The one-hot ground truth, same shape as "pred".31* @li target: A Tensor of type int32. The one-hot ground truth, same shape as "pred".
32* @li weight: An optional Tensor. Must be one of the following types: float16, float32.32* @li weight: An optional Tensor. Must be one of the following types: float16, float32.
Mmatmul/batch_mat_mul_v3/op_graph/batch_mat_mul_v3_proto.h+2-2
@@ -28,7 +28,7 @@ namespace ge {
28* float32, bfloat16. 2D-6D. Has format [ND, NHWC, NCHW].28* float32, bfloat16. 2D-6D. Has format [ND, NHWC, NCHW].
29* @li bias: A optional Tensor. Must be one of the following types: float16,29* @li bias: A optional Tensor. Must be one of the following types: float16,
30* float32, bfloat16. Has format [ND, NHWC, NCHW],30* float32, bfloat16. Has format [ND, NHWC, NCHW],
31-* bfloat16 is only supported in Ascend 950 AI processor31+* bfloat16 is only supported in Ascend 950PR&950DT products
32* @li offset_w: A optional Tensor. Must be one of the following types:32* @li offset_w: A optional Tensor. Must be one of the following types:
33* int8, int4. Has format [ND, NHWC, NCHW].33* int8, int4. Has format [ND, NHWC, NCHW].
34 34 
@@ -59,4 +59,4 @@ REG_OP(BatchMatMulV3)
59 .OP_END_FACTORY_REG(BatchMatMulV3)59 .OP_END_FACTORY_REG(BatchMatMulV3)
60} // namespace ge60} // namespace ge
61 61 
62-#endif // OPS_MATMUL_BATCH_MAT_MUL_V3_PROTO_H_62+#endif // OPS_MATMUL_BATCH_MAT_MUL_V3_PROTO_H_
Mmatmul/fused_mat_mul/op_graph/fused_mat_mul_proto.h+1-1
@@ -27,7 +27,7 @@ namespace ge {
27 * @li x2: A matrix tensor. Must be one of the following types: float16, bfloat16, float32.27 * @li x2: A matrix tensor. Must be one of the following types: float16, bfloat16, float32.
28 * float32 is supported for the the following fused_op_types: "","relu","add","mul".28 * float32 is supported for the the following fused_op_types: "","relu","add","mul".
29 * @li bias: An optional tensor. 1D. Must be one of the following types: float16, bfloat16, float32.29 * @li bias: An optional tensor. 1D. Must be one of the following types: float16, bfloat16, float32.
30- * bfloat16 is only supported in Ascend 950 AI processor. For "quant" and "relu_quant", bias only supports float16.30+ * bfloat16 is only supported in Ascend 950PR&950DT products. For "quant" and "relu_quant", bias only supports float16.
31 * @li x3: An Optional tensor. For "add", "mul" and "scale_add", x3 is the fused matrix input. For "quant" and31 * @li x3: An Optional tensor. For "add", "mul" and "scale_add", x3 is the fused matrix input. For "quant" and
32 * "relu_quant", x3 must be a uint64 tensor with shape [1], carrying the encoded logical quantization parameter.32 * "relu_quant", x3 must be a uint64 tensor with shape [1], carrying the encoded logical quantization parameter.
33 *33 *
Mmatmul/mat_mul_v3/op_graph/mat_mul_v3_proto.h+2-2
@@ -28,7 +28,7 @@ namespace ge {
28* float16, float32. Has format [ND, NHWC, NCHW, NZ].28* float16, float32. Has format [ND, NHWC, NCHW, NZ].
29* @li bias: A 1D Tensor. Must be one of the following types: float32,29* @li bias: A 1D Tensor. Must be one of the following types: float32,
30* float16, float32. Has format [ND, NHWC, NCHW]. \n30* float16, float32. Has format [ND, NHWC, NCHW]. \n
31-* bfloat16 is only supported in Ascend 950 AI processor31+* bfloat16 is only supported in Ascend 950PR&950DT products
32* @li offset_w: An optional matrix Tensor. 2D. Must be one of the following32* @li offset_w: An optional matrix Tensor. 2D. Must be one of the following
33* types: int8, int4.33* types: int8, int4.
34 34 
@@ -62,4 +62,4 @@ REG_OP(MatMulV3)
62 .OP_END_FACTORY_REG(MatMulV3)62 .OP_END_FACTORY_REG(MatMulV3)
63} // namespace ge63} // namespace ge
64 64 
65-#endif // OPS_MATMUL_MAT_MUL_V3_PROTO_H_65+#endif // OPS_MATMUL_MAT_MUL_V3_PROTO_H_
Mmatmul/quant_batch_matmul_v3/op_graph/quant_batch_matmul_v3_proto.h+10-10
@@ -116,14 +116,14 @@ values in x1 along the k-dimension. \n
116* @li The shape of bias should be 1D when the shape of out is 2D, 4D, 5D or 6D, and the shape of bias should be 1D or 3D116* @li The shape of bias should be 1D when the shape of out is 2D, 4D, 5D or 6D, and the shape of bias should be 1D or 3D
117* when the out shape is 3D.117* when the out shape is 3D.
118* @li The size of the last dimension of x1 and x2 cannot exceed 65535 only on the following computing platforms:118* @li The size of the last dimension of x1 and x2 cannot exceed 65535 only on the following computing platforms:
119-* Atlas A2 Training Series Product/Atlas A2 Inference Series Product and119+* Atlas A2 products and
120-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product.120+* Atlas A3 products.
121* The last dimension of x1 refers to m when transpose_x1 is true or k when transpose_x1 is false.121* The last dimension of x1 refers to m when transpose_x1 is true or k when transpose_x1 is false.
122* The last dimension of x2 refers to k when transpose_x2 is true or n when transpose_x2 is false.122* The last dimension of x2 refers to k when transpose_x2 is true or n when transpose_x2 is false.
123* @li If input type of x1 and x2 is int4, transpose_x1 should be false, the size of the last dimension of x1 or x2123* @li If input type of x1 and x2 is int4, transpose_x1 should be false, the size of the last dimension of x1 or x2
124should124should
125* be an even number.125* be an even number.
126-* @li On the <term>Ascend 950PR&950DT系列产品</term> platforms, when x2 is ND format, the output is an empty tensor if126+* @li On the Ascend 950PR&950DT products, when x2 is ND format, the output is an empty tensor if
127input x1 has m=0 or x2 has n=0.127input x1 has m=0 or x2 has n=0.
128* When x2 is NZ format, the output is an empty tensor if input x1 has m=0.128* When x2 is NZ format, the output is an empty tensor if input x1 has m=0.
129* In all other cases, inputs does not support tensor with dimension size 0.129* In all other cases, inputs does not support tensor with dimension size 0.
@@ -140,21 +140,21 @@ float8_e8m0:
140* - transpose_x1 must be false.140* - transpose_x1 must be false.
141* - when transpose_x2 is false, n must be greater than 2.141* - when transpose_x2 is false, n must be greater than 2.
142* - when transpose_x2 is true, n must be greater than 1.142* - when transpose_x2 is true, n must be greater than 1.
143-* @li Only weight supports ND and NZ format on Ascend 950 AI Processor. All other inputs and outputs only support ND143+* @li Only weight supports ND and NZ format on Ascend 950PR&950DT products. All other inputs and outputs only support ND
144format.144format.
145* @li When x2 is NZ format and input type of x1 and x2 is hifloat8 or float8_e4m3fn, static quantization supports145* @li When x2 is NZ format and input type of x1 and x2 is hifloat8 or float8_e4m3fn, static quantization supports
146scale uint64/int64, and dynamic quantization supports scale and pertoken_scale both float32.146scale uint64/int64, and dynamic quantization supports scale and pertoken_scale both float32.
147* @li The following are the supported data type combinations by platform.147* @li The following are the supported data type combinations by platform.
148 148 
149-* - Atlas Inference Series Product:149+* - Atlas inference products:
150*\n150*\n
151| x1 | x2 | scale | offset | bias | pertoken | out |151| x1 | x2 | scale | offset | bias | pertoken | out |
152| :------: | :------: | :----------: | :-----------: | :-----------: | :------: | :------: |152| :------: | :------: | :----------: | :-----------: | :-----------: | :------: | :------: |
153| int8 | int8 | uint64/int64 | null | null/int32 | null | float16 |153| int8 | int8 | uint64/int64 | null | null/int32 | null | float16 |
154| int8 | int8 | uint64/int64 | null/float32 | null/int32 | null | int8 |154| int8 | int8 | uint64/int64 | null/float32 | null/int32 | null | int8 |
155*\n155*\n
156-* - Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or156+* - Atlas A2 products or
157-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product:157+* Atlas A3 products:
158*\n158*\n
159| x1 | x2 | scale | offset | bias | pertoken | out |159| x1 | x2 | scale | offset | bias | pertoken | out |
160| :------: | :------: | :--------------: | :-----------: | :-------------------------: | :----------: | :------: |160| :------: | :------: | :--------------: | :-----------: | :-------------------------: | :----------: | :------: |
@@ -167,7 +167,7 @@ scale uint64/int64, and dynamic quantization supports scale and pertoken_scale b
167| int4 | int4 | float32/bfloat16 | null | null/int32/bfloat16/float32 | float32 | bfloat16 |167| int4 | int4 | float32/bfloat16 | null | null/int32/bfloat16/float32 | float32 | bfloat16 |
168| int4 | int4 | float32 | null | null/int32/float16/float32 | float32 | float16 |168| int4 | int4 | float32 | null | null/int32/float16/float32 | float32 | float16 |
169*\n169*\n
170-* - Ascend 950 AI Processor:170+* - Ascend 950PR&950DT products:
171*\n171*\n
172| x1 | x2 | scale | offset | bias | pertoken |172| x1 | x2 | scale | offset | bias | pertoken |
173out | | :-----------------------: | :-----------------------: | :------------------:173out | | :-----------------------: | :-----------------------: | :------------------:
@@ -188,7 +188,7 @@ float16/bfloat16/float32 | | float4_e2m1 | float4_e2
188null/float32 | float8_e8m0 | float16/bfloat16/float32 | | int8 | int8188null/float32 | float8_e8m0 | float16/bfloat16/float32 | | int8 | int8
189| float32/bfloat16 | null | null/int32 | null | int32 |189| float32/bfloat16 | null | null/int32 | null | int32 |
190*\n190*\n
191-* - Ascend 950 AI Processor, supported data type and quant mode combinations:191+* - Ascend 950PR&950DT products, supported data type and quant mode combinations:
192*\n192*\n
193pertensor-perchannel && pertensor-pertensor:193pertensor-perchannel && pertensor-pertensor:
194*\n194*\n
@@ -223,7 +223,7 @@ mx quant：
223 | float8_e4m3fn/float8_e5m2 | float8_e4m3fn/float8_e5m2 | float8_e8m0 | float8_e8m0 |223 | float8_e4m3fn/float8_e5m2 | float8_e4m3fn/float8_e5m2 | float8_e8m0 | float8_e8m0 |
224 | float4_e2m1 | float4_e2m1 | float8_e8m0 | float8_e8m0 |224 | float4_e2m1 | float4_e2m1 | float8_e8m0 | float8_e8m0 |
225*\n225*\n
226-* - Ascend 950 AI Processor with group_sizes scenarios, supported data type and shapes combinations:226+* - Ascend 950PR&950DT products with group_sizes scenarios, supported data type and shapes combinations:
227*\n227*\n
228| quantization | x1 type | scale type | x1 shape | x2 shape | scale shape |228| quantization | x1 type | scale type | x1 shape | x2 shape | scale shape |
229pertoken shape | group_size |229pertoken shape | group_size |
Mmatmul/quant_matmul_activation_quant/op_graph/quant_matmul_activation_quant_proto.h+3-3
@@ -110,12 +110,12 @@ even.
110 * - scale_alg only supports 0 and 2. When scale_alg is 2, dst_type_max supports 0.0 and 6.0-12.0.110 * - scale_alg only supports 0 and 2. When scale_alg is 2, dst_type_max supports 0.0 and 6.0-12.0.
111 * - round_mode supports "rint", "floor", "round".111 * - round_mode supports "rint", "floor", "round".
112 * - When x2 is NZ format, x1 does not support transpose, i.e. transpose_x1 cannot be true.112 * - When x2 is NZ format, x1 does not support transpose, i.e. transpose_x1 cannot be true.
113- * @li Only weight supports ND and NZ format on Ascend 950 AI Processor. All other inputs and outputs only support113+ * @li Only weight supports ND and NZ format on Ascend 950PR&950DT products. All other inputs and outputs only support
114ND114ND
115format.115format.
116* @li The following are the supported data type combinations by platform.116* @li The following are the supported data type combinations by platform.
117 117 
118-* - Ascend 950 AI Processor:118+* - Ascend 950PR&950DT products:
119*\n119*\n
120| x1 | x2 | bias | x1_scale | x2_scale | out |120| x1 | x2 | bias | x1_scale | x2_scale | out |
121y_scale | | :-----------------------: | :-----------: | :----------: | :---------: | :---------: |121y_scale | | :-----------------------: | :-----------: | :----------: | :---------: | :---------: |
@@ -127,7 +127,7 @@ float8_e4m3fn | null/float32 | float8_e8m0 | float8_e8m0 | float8_e5m2
127float4_e2m1 | null/float32 | float8_e8m0 | float8_e8m0 | float4_e2m1 | float8_e8m0 |127float4_e2m1 | null/float32 | float8_e8m0 | float8_e8m0 | float4_e2m1 | float8_e8m0 |
128*\n128*\n
129 129 
130-* - Ascend 950 AI Processor with group_sizes scenarios, supported data type and shapes combinations:130+* - Ascend 950PR&950DT products with group_sizes scenarios, supported data type and shapes combinations:
131*\n131*\n
132| quantization | x1 type | x1_scale type | x1 shape | x2 shape | x2_scale132| quantization | x1 type | x1_scale type | x1 shape | x2 shape | x2_scale
133shape | x1_scale shape | group_size |133shape | x1_scale shape | group_size |
Mmatmul/quant_matmul_reduce_sum/op_graph/quant_matmul_reduce_sum_proto.h+1-1
@@ -48,7 +48,7 @@ namespace ge {
48* @par Outputs:48* @par Outputs:
49* y: A matrix tensor. The data type is bfloat16. The format supports ND. \n49* y: A matrix tensor. The data type is bfloat16. The format supports ND. \n
50 50 
51-* Atlas A2 Trainging Series Product/Atlas 800I A2 Inference Product or Atlas A3 Training Series Product: \n51+* Atlas A2 products or Atlas A3 training products: \n
52* | x1 | x2 | x1Scale | x2Scale | yScale | x1Offset | x2Offset | yOffset | bias | out |52* | x1 | x2 | x1Scale | x2Scale | yScale | x1Offset | x2Offset | yOffset | bias | out |
53* |------|------|----------|----------|--------|----------|----------|---------|-------|---------|53* |------|------|----------|----------|--------|----------|----------|---------|-------|---------|
54* | INT8 | INT8 | FLOAT32 | BFLOAT16 | null | null | null | null | null |BFLOAT16 |54* | INT8 | INT8 | FLOAT32 | BFLOAT16 | null | null | null | null | null |BFLOAT16 |
Mmatmul/rotate_quant/op_graph/rotate_quant_proto.h+4-4
@@ -47,16 +47,16 @@ namespace ge {
47* @li trans: An optional bool. Specifies whether to transpose. Defaults to false. \n47* @li trans: An optional bool. Specifies whether to transpose. Defaults to false. \n
48 48 
49* @par Constraints:49* @par Constraints:
50-* Atlas A3 supports per-token dynamic quantization.Atlas A5 supports per-group dynamic MX quantization.50+* Atlas A3 supports per-token dynamic quantization. Ascend 950PR&950DT products support per-group dynamic MX
51-* Atlas A3 Training Series Products/Atlas A3 Inference Series Products, Atlas A2 Training Series Products/Atlas A251+quantization.
52-Inference Series Products, Atlas 950 Series Products: \n52+* Atlas A3 products, Atlas A2 products, Ascend 950PR&950DT products: \n
53* - x shape is [M, N], rot shape is [K, K]. rot must be a square matrix.53* - x shape is [M, N], rot shape is [K, K]. rot must be a square matrix.
54* - N must be a multiple of K, and N must be divisible by 8.54* - N must be a multiple of K, and N must be divisible by 8.
55* - x and rot must have the same data type.55* - x and rot must have the same data type.
56* - scale output shape must be [M].56* - scale output shape must be [M].
57* - N range: [128, 16000], K range: [16, 1024]. \n57* - N range: [128, 16000], K range: [16, 1024]. \n
58 58 
59-* Atlas A5 Series Products:59+* Ascend 950PR&950DT products:
60* - x is 1-D to 7-D, with the last dimension being N.60* - x is 1-D to 7-D, with the last dimension being N.
61* - rot shape is [K, K] or [blockNum, K, K], where blockNum = K/N, and N must divide K.61* - rot shape is [K, K] or [blockNum, K, K], where blockNum = K/N, and N must divide K.
62* - x and rot must have the same data type.62* - x and rot must have the same data type.
Mmatmul/weight_quant_batch_matmul_v2/op_graph/weight_quant_batch_matmul_v2_proto.h+4-5
@@ -28,7 +28,7 @@ namespace ge {
28* @li weight: A matrix tensor of quantized weight. Shape supports (n,k)/(k,n),28* @li weight: A matrix tensor of quantized weight. Shape supports (n,k)/(k,n),
29* Format supports ND/NZ. The type support int8, int4, int32, float8_e4m3fn, hifloat8, float4_e2m1. \n29* Format supports ND/NZ. The type support int8, int4, int32, float8_e4m3fn, hifloat8, float4_e2m1. \n
30* Format must be ND when the type is float8_e4m3fn or hifloat8.30* Format must be ND when the type is float8_e4m3fn or hifloat8.
31-* For Ascend 950 AI Processor, transpose_weight must be false when format is NZ. \n31+* For Ascend 950PR&950DT products, transpose_weight must be false when format is NZ. \n
32* The k, n value must be at least 1.32* The k, n value must be at least 1.
33* The k value must be even when type is int4/float4_e2m1 and transpose_weight33* The k value must be even when type is int4/float4_e2m1 and transpose_weight
34* is true, and the n value must be even when type is int4/float4_e2m1 and transpose_weight34* is true, and the n value must be even when type is int4/float4_e2m1 and transpose_weight
@@ -91,10 +91,9 @@ namespace ge {
91* The type should be the same with x when quant_scale not exits.91* The type should be the same with x when quant_scale not exits.
92 92 
93* @attention Constraints:93* @attention Constraints:
94-* @li It is not recommended to use weight NZ format on Atlas A2 Training Series Product/Atlas 800I A2 Inference94+* @li It is not recommended to use weight NZ format on Atlas A2 products,
95-Product/A200I A2 Box Heterogeneous Component,
96* because its performance may not be better than ND format.95* because its performance may not be better than ND format.
97-* @li All of these conditions must be met on Atlas Inference Series Product: weight type is int8,96+* @li All of these conditions must be met on Atlas inference products: weight type is int8,
98* weight format is NZ, transpose_weight is true, x type is float16, antiquant_scale only support the per_channel mode,97* weight format is NZ, transpose_weight is true, x type is float16, antiquant_scale only support the per_channel mode,
99* quant_scale not exists, quant_offset not exists, antiquant_group_size is 0.98* quant_scale not exists, quant_offset not exists, antiquant_group_size is 0.
100* @li Per_channel mode: To improve performance, it is recommended to use the99* @li Per_channel mode: To improve performance, it is recommended to use the
@@ -118,4 +117,4 @@ REG_OP(WeightQuantBatchMatmulV2)
118 .OP_END_FACTORY_REG(WeightQuantBatchMatmulV2)117 .OP_END_FACTORY_REG(WeightQuantBatchMatmulV2)
119} // namespace ge118} // namespace ge
120 119 
121-#endif // OPS_MATMUL_WEIGHT_QUANT_BATCH_MATMUL_V2_PROTO_H_120+#endif // OPS_MATMUL_WEIGHT_QUANT_BATCH_MATMUL_V2_PROTO_H_
Mnorm/batch_norm3_d/op_graph/batch_norm3d_proto.h+0-3
@@ -58,9 +58,6 @@ Specifies the variance of "x" for gradient computation. Pass "None" to skip this
58*@attention Constraints:58*@attention Constraints:
59*@li If the operation is used for inference and outputs "reserve_space_1" and "reserve_space_2" are available,59*@li If the operation is used for inference and outputs "reserve_space_1" and "reserve_space_2" are available,
60then "reserve_space_1" has the same value as "mean" and "reserve_space_2" has the same value as "variance".60then "reserve_space_1" has the same value as "mean" and "reserve_space_2" has the same value as "variance".
61-*@li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1‰ due to the square root instruction .
62-\n
63- 
64*@par Third-party framework compatibility61*@par Third-party framework compatibility
65*@li Compatible with the TensorFlow operator fused_batch_norm.62*@li Compatible with the TensorFlow operator fused_batch_norm.
66*@li Compatible with the TensorFlow operator fused_batch_norm_v2.63*@li Compatible with the TensorFlow operator fused_batch_norm_v2.
Mnorm/batch_norm_ext2/op_graph/batch_norm_ext2_proto.h+0-2
@@ -50,8 +50,6 @@ is used for training or inference. Defaults to "True" . \n
50*@attention Constraints:50*@attention Constraints:
51*@li If the operation is used for inference, then output "reserve_space_1"51*@li If the operation is used for inference, then output "reserve_space_1"
52has the same value as "mean" and output "reserve_space_2" has the same value as "variance".52has the same value as "mean" and output "reserve_space_2" has the same value as "variance".
53-*@li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1‰ due to the square root instruction .
54- 
55*@par Third-party framework compatibility53*@par Third-party framework compatibility
56* Compatible with the TensorFlow operator fused_batch_norm_v2.54* Compatible with the TensorFlow operator fused_batch_norm_v2.
57*/55*/
Mnorm/bn3_d_training_update/op_graph/bn3_d_training_update_proto.h+0-2
@@ -51,8 +51,6 @@ channel.
51* @li This operator is a BatchNorm fusion operator for updating the moving51* @li This operator is a BatchNorm fusion operator for updating the moving
52* averages for training.52* averages for training.
53* This operator is used in conjunction with BN3DTrainingUpdate.53* This operator is used in conjunction with BN3DTrainingUpdate.
54-* @li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the square
55-* root instruction.
56*/54*/
57#ifndef OPS_PROTO_DEF_BN3DTRAININGUPDATE55#ifndef OPS_PROTO_DEF_BN3DTRAININGUPDATE
58#define OPS_PROTO_DEF_BN3DTRAININGUPDATE56#define OPS_PROTO_DEF_BN3DTRAININGUPDATE
Mnorm/bn_infer/op_graph/bn_infer_proto.h+0-3
@@ -32,9 +32,6 @@ namespace ge {
32 * @par Outputs:32 * @par Outputs:
33 * y: A tensor of type float16 or float32 or bfloat16 for the normalized "x", with the same format as x. \n33 * y: A tensor of type float16 or float32 or bfloat16 for the normalized "x", with the same format as x. \n
34 *34 *
35- * @attention Constraints:
36- * For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the
37- * square root instruction.
38 */35 */
39#ifndef OPS_PROTO_DEF_BNINFER36#ifndef OPS_PROTO_DEF_BNINFER
40#define OPS_PROTO_DEF_BNINFER37#define OPS_PROTO_DEF_BNINFER
Mnorm/bn_inference/op_graph/bn_inference_proto.h+7-7
@@ -22,23 +22,23 @@ namespace ge {
22 * @brief Performs inference batch normalization on a 4-D or 5-D tensor.22 * @brief Performs inference batch normalization on a 4-D or 5-D tensor.
23 *23 *
24 * @par Inputs:24 * @par Inputs:
25- * @li x: The feature tensor. On Ascend 950PR/Ascend 950DT, NCHW/NHWC use four dimensions, NCDHW/NDHWC use five25+ * @li x: The feature tensor. On Ascend 950PR&950DT products, NCHW/NHWC use four dimensions, NCDHW/NDHWC use five
26 * dimensions, and ND uses four or five dimensions. NCHW/NCDHW use channel axis 1, NHWC/NDHWC use the last axis,26 * dimensions, and ND uses four or five dimensions. NCHW/NCDHW use channel axis 1, NHWC/NDHWC use the last axis,
27 * and ND storage follows its public origin format or uses axis 1 when the origin format is ND.27 * and ND storage follows its public origin format or uses axis 1 when the origin format is ND.
28- * @li mean: A one-dimensional tensor of length C. On Ascend 950PR/Ascend 950DT, it is the original mean when mode is28+ * @li mean: A one-dimensional tensor of length C. On Ascend 950PR&950DT products, it is the original mean when mode is
29 * nonzero and the pre-folded alpha when mode is 0.29 * nonzero and the pre-folded alpha when mode is 0.
30- * @li variance: A one-dimensional tensor of length C. On Ascend 950PR/Ascend 950DT, it is the original variance when30+ * @li variance: A one-dimensional tensor of length C. On Ascend 950PR&950DT products, it is the original variance when
31 * mode is nonzero and the pre-folded beta when mode is 0.31 * mode is nonzero and the pre-folded beta when mode is 0.
32- * @li momentum: The momentum tensor. On Ascend 950PR/Ascend 950DT, its shape is [], [1], or [C]. Its first element is32+ * @li momentum: The momentum tensor. On Ascend 950PR&950DT products, its shape is [], [1], or [C]. Its first element is
33 * used only when mode is nonzero and both optional inputs are absent.33 * used only when mode is nonzero and both optional inputs are absent.
34- * @li scale: Optional one-dimensional tensor of length C. On Ascend 950PR/Ascend 950DT, scale-only is supported.34+ * @li scale: Optional one-dimensional tensor of length C. On Ascend 950PR&950DT products, scale-only is supported.
35- * @li offset: Optional one-dimensional tensor of length C. On Ascend 950PR/Ascend 950DT, offset-only is supported35+ * @li offset: Optional one-dimensional tensor of length C. On Ascend 950PR&950DT products, offset-only is supported
36 * when mode is nonzero; mode 0 requires scale when offset is present.36 * when mode is nonzero; mode 0 requires scale when offset is present.
37 *37 *
38 * @par Attributes:38 * @par Attributes:
39 * @li epsilon: An optional float. The default value is 1e-5.39 * @li epsilon: An optional float. The default value is 1e-5.
40 * @li use_global_stats: An optional bool. The default value is true.40 * @li use_global_stats: An optional bool. The default value is true.
41- * @li mode: An optional integer. The default value is 1. On Ascend 950PR/Ascend 950DT, 0 selects the pre-folded41+ * @li mode: An optional integer. The default value is 1. On Ascend 950PR&950DT products, 0 selects the pre-folded
42 * expression and every nonzero integer selects complete BNInference behavior. On other products, mode does not42 * expression and every nonzero integer selects complete BNInference behavior. On other products, mode does not
43 * change the complete BNInference behavior.43 * change the complete BNInference behavior.
44 *44 *
Mnorm/bn_training_update_v2/op_graph/bn_training_update_v2_proto.h+0-2
@@ -54,8 +54,6 @@ format as "x". \n
54 54 
55* @attention Constraints:55* @attention Constraints:
56* @li This operator is used in conjunction with BNTrainingReduce.56* @li This operator is used in conjunction with BNTrainingReduce.
57-* @li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to
58-* the square root instruction.
59*/57*/
60#ifndef OPS_PROTO_DEF_BNTRAININGUPDATEV258#ifndef OPS_PROTO_DEF_BNTRAININGUPDATEV2
61#define OPS_PROTO_DEF_BNTRAININGUPDATEV259#define OPS_PROTO_DEF_BNTRAININGUPDATEV2
Mnorm/bn_training_update_v3/op_graph/bn_training_update_v3_proto.h+3-5
@@ -58,14 +58,12 @@ format as "x".
58 58 
59* @attention Constraints:59* @attention Constraints:
60* @li This operator is used in conjunction with BNTrainingReduce.60* @li This operator is used in conjunction with BNTrainingReduce.
61-* @li For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to61+* @li The 4D NHWC/NCHW description above applies to the Atlas A2 products. For Ascend 950PR&950DT products: "x" supports
62-* the square root instruction.62+ND (or NCHW tag, layout-normalized) with rank 2~8 ([N, C, R...],
63-* @li The 4D NHWC/NCHW description above applies to the A2-series products. For Ascend 950PR/
64-* Ascend 950DT: "x" supports ND (or NCHW tag, layout-normalized) with rank 2~8 ([N, C, R...],
65* dim0=N, dim1=C); NHWC/NC1HWC0/NDC1HWC0 origin formats are rejected by tiling. The statistics63* dim0=N, dim1=C); NHWC/NC1HWC0/NDC1HWC0 origin formats are rejected by tiling. The statistics
66* inputs/outputs hold element count C (recommended shape [C]; output shapes follow "scale").64* inputs/outputs hold element count C (recommended shape [C]; output shapes follow "scale").
67* Optional A2 attributes before_split_ori_shape/before_split_ori_format (BN FFTS scenario) are65* Optional A2 attributes before_split_ori_shape/before_split_ori_format (BN FFTS scenario) are
68-* not registered on Ascend 950PR/Ascend 950DT and have no effect there.66+* not registered on Ascend 950PR&950DT products and have no effect there.
69*/67*/
70#ifndef OPS_PROTO_DEF_BNTRAININGUPDATEV368#ifndef OPS_PROTO_DEF_BNTRAININGUPDATEV3
71#define OPS_PROTO_DEF_BNTRAININGUPDATEV369#define OPS_PROTO_DEF_BNTRAININGUPDATEV3
Mnorm/group_norm_silu/op_graph/group_norm_silu_proto.h+2-2
@@ -40,10 +40,10 @@ namespace ge {
40* @li y: A ND Tensor of type bfloat16/float16/float32 for the normalized "x".40* @li y: A ND Tensor of type bfloat16/float16/float32 for the normalized "x".
41* @li mean: A Tensor of type bfloat16/float16/float32.41* @li mean: A Tensor of type bfloat16/float16/float32.
42* Must be 1D. Specifies the mean of "x".42* Must be 1D. Specifies the mean of "x".
43-* Ascend 950 AI Processor: The data type must be consistent with that of gamma and beta.43+* Ascend 950PR&950DT products: The data type must be consistent with that of gamma and beta.
44* @li rstd: A Tensor of type bfloat16/float16/float32.44* @li rstd: A Tensor of type bfloat16/float16/float32.
45* Must be 1D. Specifies the rstd of "x".45* Must be 1D. Specifies the rstd of "x".
46-* Ascend 950 AI Processor: The data type must be consistent with that of gamma and beta.46+* Ascend 950PR&950DT products: The data type must be consistent with that of gamma and beta.
47 47 
48* @par Third-party framework compatibility48* @par Third-party framework compatibility
49* @li Compatible with the PyTorch operator GroupNorm and Silu.49* @li Compatible with the PyTorch operator GroupNorm and Silu.
Mnorm/in_infer_v2/op_graph/in_infer_v2_proto.h+0-2
@@ -41,8 +41,6 @@ Defaults to "0.00001" . \n
41*@li batch_mean: A Tensor of type float32 for the result mean.41*@li batch_mean: A Tensor of type float32 for the result mean.
42*@li batch_variance: A Tensor of type float32 for the result variance . \n42*@li batch_variance: A Tensor of type float32 for the result variance . \n
43 43 
44-*@attention Constraints:
45-*For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 0.001 due to the square root instruction.
46*/44*/
47#ifndef OPS_PROTO_DEF_ININFERV245#ifndef OPS_PROTO_DEF_ININFERV2
48#define OPS_PROTO_DEF_ININFERV246#define OPS_PROTO_DEF_ININFERV2
Mnorm/layer_norm_grad_v3/op_graph/layer_norm_grad_v3_proto.h+1-1
@@ -75,7 +75,7 @@ derivative.
75* The output_mask[0] is false, means pd_x is meaningless.75* The output_mask[0] is false, means pd_x is meaningless.
76* The output_mask[1] is false, means pd_gamma is meaningless.76* The output_mask[1] is false, means pd_gamma is meaningless.
77* The output_mask[2] is false, means pd_beta is meaningless.77* The output_mask[2] is false, means pd_beta is meaningless.
78-* Except in Ascend 950 AI Processor, the output_mask just supports {true, true, true} in other AI Processor.78+* Except in Ascend 950PR&950DT products, the output_mask just supports {true, true, true} in other AI Processor.
79 79 
80*@par Restrictions:80*@par Restrictions:
81*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.81*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
Mpooling/adaptive_avg_pool3d/op_graph/adaptive_avg_pool3d_proto.h+3-3
@@ -27,14 +27,14 @@ namespace ge {
27* One input, including:27* One input, including:
28* @li x: A Tensor. Must be one of the following data types:28* @li x: A Tensor. Must be one of the following data types:
29* float16, bfloat16, float32. \n29* float16, bfloat16, float32. \n
30-* For Ascend 950PR/Ascend 950DT: Support Format: [NDHWC, NCDHW].30+* For Ascend 950PR&950DT products: Support Format: [NDHWC, NCDHW].
31 31 
32* @par Attributes:32* @par Attributes:
33* @li output_size: A required list of 3 ints33* @li output_size: A required list of 3 ints
34* specifying the size (D,H,W) of the output tensor. \n34* specifying the size (D,H,W) of the output tensor. \n
35* @li data_format: An optional string, Specify the data format of the input and35* @li data_format: An optional string, Specify the data format of the input and
36* output data. either "NDHWC"(default), or "NCDHW" . \n36* output data. either "NDHWC"(default), or "NCDHW" . \n
37-* For Ascend 950PR/Ascend 950DT: Supported "NDHWC", and "NCDHW".37+* For Ascend 950PR&950DT products: Supported "NDHWC", and "NCDHW".
38 38 
39* @par Outputs:39* @par Outputs:
40* @li y: A Tensor. Has the same data type and the same format as "x" \n40* @li y: A Tensor. Has the same data type and the same format as "x" \n
@@ -51,4 +51,4 @@ REG_OP(AdaptiveAvgPool3d)
51 51 
52} // namespace ge52} // namespace ge
53 53 
54-#endif // OPS_POOLING_ADAPTIVE_AVG_POOL3D_PROTO_H_54+#endif // OPS_POOLING_ADAPTIVE_AVG_POOL3D_PROTO_H_
Mpooling/adaptive_avg_pool3d_grad/op_graph/adaptive_avg_pool3d_grad_proto.h+1-1
@@ -37,7 +37,7 @@ namespace ge {
37* @par Attributes37* @par Attributes
38* data_format: An optional string, Specify the data format of the input and38* data_format: An optional string, Specify the data format of the input and
39* output data. With the default format "NDHWC". \n39* output data. With the default format "NDHWC". \n
40-* For Ascend 950PR/Ascend 950DT, both "NDHWC" and "NCDHW" are supported. All other platforms support only "NDHWC".40+* For Ascend 950PR&950DT products, both "NDHWC" and "NCDHW" are supported. All other platforms support only "NDHWC".
41 41 
42* @par Third-party framework compatibility42* @par Third-party framework compatibility
43* Compatible with the PyTorch operator AdaptiveAvgPool3dGrad.43* Compatible with the PyTorch operator AdaptiveAvgPool3dGrad.
Mpooling/avg_pool/op_graph/avg_pool_proto.h+7-7
@@ -43,19 +43,19 @@ namespace ge {
43* @attention Constraints:43* @attention Constraints:
44* @li This operator applies only to a TensorFlow network.44* @li This operator applies only to a TensorFlow network.
45* @li Only single input and single output are supported.45* @li Only single input and single output are supported.
46-* @li For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,46+* @li For Atlas training products, Atlas A2 products,
47-* Atlas A3 Training Series Product: "ksize_H" and "ksize_W" are positive integers within the range [1, 255].47+* Atlas A3 training products: "ksize_H" and "ksize_W" are positive integers within the range [1, 255].
48* ksize_H * ksize_W < 256. \n48* ksize_H * ksize_W < 256. \n
49-* For Ascend 950 AI Processor: The ksize of the H and W dimensions should be greater than 0.49+* For Ascend 950PR&950DT products: The ksize of the H and W dimensions should be greater than 0.
50-* @li For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,50+* @li For Atlas training products, Atlas A2 products,
51-* Atlas A3 Training Series Product: the values of "strides_h" and "strides_w" are positive integers within51+* Atlas A3 training products: the values of "strides_h" and "strides_w" are positive integers within
52* the range [1, 63]. \n52* the range [1, 63]. \n
53-* For Ascend 950 AI Processor: The stride of the H and W dimensions should be greater than 0.53+* For Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0.
54* @li When the C axis is greater than 1, if points with the same H and W dimensions in x contain one INF input54* @li When the C axis is greater than 1, if points with the same H and W dimensions in x contain one INF input
55* on the C axis, the output of the INF input covered by the sliding window on this C axis is INF, and the55* on the C axis, the output of the INF input covered by the sliding window on this C axis is INF, and the
56* outputs of other C axis without INF input covered by the sliding window are Nan. If points with the same56* outputs of other C axis without INF input covered by the sliding window are Nan. If points with the same
57* H and W dimensions in x contain more than one INF input on the C axis, the outputs of all INF input data57* H and W dimensions in x contain more than one INF input on the C axis, the outputs of all INF input data
58-* covered by the sliding window on the C axis are Nan. this constraints not for Ascend 950 AI Processor.58+* covered by the sliding window on the C axis are Nan. this constraints not for Ascend 950PR&950DT products.
59* @li The output "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output59* @li The output "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output
60* shape at the H and W dimensions is calculated by below formula: \n60* shape at the H and W dimensions is calculated by below formula: \n
61* @code{.c}61* @code{.c}
Mpooling/avg_pool3_d/op_graph/avg_pool3_d_proto.h+9-9
@@ -22,8 +22,8 @@ namespace ge {
22* @brief Performs average pooling on the input.22* @brief Performs average pooling on the input.
23 23 
24* @par Inputs:24* @par Inputs:
25-* x: A 5D tensor. Supported type:float16. Additional support for float32, bfloat16 in Ascend 950 AI Processor.25+* x: A 5D tensor. Supported type:float16. Additional support for float32, bfloat16 in Ascend 950PR&950DT products.
26-* Support format: NDHWC. Additional support for NCDHW in Ascend 950 AI Processor. \n26+* Support format: NDHWC. Additional support for NCDHW in Ascend 950PR&950DT products. \n
27 27 
28* @par Attributes:28* @par Attributes:
29* @li ksize: List of ints that has length 1, 3 or 5. The size of the window29* @li ksize: List of ints that has length 1, 3 or 5. The size of the window
@@ -39,20 +39,20 @@ namespace ge {
39* @li divisor_override: if specified, it will be used as divisor, otherwise39* @li divisor_override: if specified, it will be used as divisor, otherwise
40* size of the pooling region will be used. Default value is 0.40* size of the pooling region will be used. Default value is 0.
41* @li data_format: An optional string, format of input data. It supports "NDHWC"(default).41* @li data_format: An optional string, format of input data. It supports "NDHWC"(default).
42-* Additional support for "NCDHW" in Ascend 950 AI Processor. \n42+* Additional support for "NCDHW" in Ascend 950PR&950DT products. \n
43 43 
44* @par Outputs:44* @par Outputs:
45* y: The average pooled output tensor. \n45* y: The average pooled output tensor. \n
46 46 
47* @attention Constraints:47* @attention Constraints:
48* @li The ksize of the D, H and W dimensions should be greater than 0.48* @li The ksize of the D, H and W dimensions should be greater than 0.
49-* For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,49+* For Atlas training products, Atlas A2 products,
50-* Atlas A3 Training Series Product: The produce of the ksize in D, H and W dimensions50+* Atlas A3 training products: The produce of the ksize in D, H and W dimensions
51* should be greater than 0 and less than or equal to 255. \n51* should be greater than 0 and less than or equal to 255. \n
52-* @li For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,52+* @li For Atlas training products, Atlas A2 products,
53-* Atlas A3 Training Series Product: The stride of the D, H and W dimensions should be greater than 0 and53+* Atlas A3 training products: The stride of the D, H and W dimensions should be greater than 0 and
54* smaller than 64. \n54* smaller than 64. \n
55-* For Ascend 950 AI Processor: The stride of the D, H and W dimensions should be greater than 0.55+* For Ascend 950PR&950DT products: The stride of the D, H and W dimensions should be greater than 0.
56* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output56* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output
57* shape at the D, H and W dimensions is calculated by below formula: \n57* shape at the D, H and W dimensions is calculated by below formula: \n
58* @code{.c}58* @code{.c}
@@ -90,4 +90,4 @@ REG_OP(AvgPool3D)
90 90 
91} // namespace ge91} // namespace ge
92 92 
93-#endif // OPS_POOLING_AVG_POOL3D_PROTO_H_93+#endif // OPS_POOLING_AVG_POOL3D_PROTO_H_
Mpooling/avg_pool3_d_grad/op_graph/avg_pool3_d_grad_proto.h+22-21
@@ -32,29 +32,30 @@ namespace ge {
32* @par Attributes:32* @par Attributes:
33* @li ksize: List of ints. Describes the size of the window for each dimension of the input tensor.33* @li ksize: List of ints. Describes the size of the window for each dimension of the input tensor.
34* Restriction: "ksize" value is in the range [1, 255] and <= "orig_input_shape" for DHW dimensions. \n34* Restriction: "ksize" value is in the range [1, 255] and <= "orig_input_shape" for DHW dimensions. \n
35-* For Atlas Inference Series Product: "ksize" length is 5. \n35+* For Atlas inference products: "ksize" length is 5. \n
36-* For Atlas Training Series Product: "ksize" length is 5. \n36+* For Atlas training products: "ksize" length is 5. \n
37-* For Atlas A2 Training Series Product/Atlas A2 Inference Series Product: "ksize"37+* For Atlas A2 products: "ksize"
38length is 3. \n38length is 3. \n
39-* For Atlas A3 Training Series Product/Atlas A3 Inference Series Product: "ksize" length is 3. \n39+* For Atlas A3 products: "ksize" length is 3. \n
40-* For Ascend950PR/Ascend950DT: "ksize" length is 1, 3 or 5, without the limit of [1, 255] and must be greater than 0. \n40+* For Ascend 950PR&950DT products: "ksize" length is 1, 3 or 5, without the limit of [1, 255] and must be greater than
41+0. \n
41* @li strides: List of ints. The stride of the sliding window for each dimension of the input tensor.42* @li strides: List of ints. The stride of the sliding window for each dimension of the input tensor.
42* Restriction: "strides" value is in the range [1, 63]. \n43* Restriction: "strides" value is in the range [1, 63]. \n
43-* For Atlas Inference Series Product: "strides" length is 5. \n44+* For Atlas inference products: "strides" length is 5. \n
44-* For Atlas Training Series Product: "strides" length is 5. \n45+* For Atlas training products: "strides" length is 5. \n
45-* For Atlas A2 Training Series Product/Atlas A2 Inference Series Product: "strides"46+* For Atlas A2 products: "strides"
46length is 3. \n47length is 3. \n
47-* For Atlas A3 Training Series Product/Atlas A3 Inference Series Product: "strides" length is 3. \n48+* For Atlas A3 products: "strides" length is 3. \n
48-* For Ascend950PR/Ascend950DT: "strides" length is 1, 3 or 5, without the limit of [1, 63] and must be greater than 0.49+* For Ascend 950PR&950DT products: "strides" length is 1, 3 or 5, without the limit of [1, 63] and must be greater than
49-\n50+0. \n
50* @li pads: List of ints, implicit zero paddings on both sides of the input.51* @li pads: List of ints, implicit zero paddings on both sides of the input.
51* Restriction: "pads" is in the range [0, ksize/2]. \n52* Restriction: "pads" is in the range [0, ksize/2]. \n
52-* For Atlas Inference Series Product: "pads" length is 6. \n53+* For Atlas inference products: "pads" length is 6. \n
53-* For Atlas Training Series Product: "pads" length is 6. \n54+* For Atlas training products: "pads" length is 6. \n
54-* For Atlas A2 Training Series Product/Atlas A2 Inference Series Product: "pads"55+* For Atlas A2 products: "pads"
55length is 3. \n56length is 3. \n
56-* For Atlas A3 Training Series Product/Atlas A3 Inference Series Product: "pads" length is 3. \n57+* For Atlas A3 products: "pads" length is 3. \n
57-* For Ascend950PR/Ascend950DT: "pads" length is 1, 3 or 6. \n58+* For Ascend 950PR&950DT products: "pads" length is 1, 3 or 6. \n
58* @li ceil_mode: An optional bool. When true, will use ceil instead of floor in the formula to59* @li ceil_mode: An optional bool. When true, will use ceil instead of floor in the formula to
59* compute the output shape. Default value false.60* compute the output shape. Default value false.
60* @li count_include_pad: An optional bool. When true, will include the zero-padding in the61* @li count_include_pad: An optional bool. When true, will include the zero-padding in the
@@ -62,12 +63,12 @@ length is 3. \n
62* @li divisor_override: An optional int, if specified, it will be used as divisor, otherwise63* @li divisor_override: An optional int, if specified, it will be used as divisor, otherwise
63* size of the pooling region will be used. Default value 0, which means this attribute does not take effect.64* size of the pooling region will be used. Default value 0, which means this attribute does not take effect.
64* @li data_format: An optional string, the format of the input "grads". Defaults to "NDHWC".65* @li data_format: An optional string, the format of the input "grads". Defaults to "NDHWC".
65-* For Atlas Inference Series Product: support "NDHWC" and "NCDHW". \n66+* For Atlas inference products: support "NDHWC" and "NCDHW". \n
66-* For Atlas Training Series Product: support "NDHWC" and "NCDHW". \n67+* For Atlas training products: support "NDHWC" and "NCDHW". \n
67-* For Atlas A2 Training Series Product/Atlas A2 Inference Series Product: only68+* For Atlas A2 products: only
68support "NDHWC". \n69support "NDHWC". \n
69-* For Atlas A3 Training Series Product/Atlas A3 Inference Series Product: only support "NDHWC". \n70+* For Atlas A3 products: only support "NDHWC". \n
70-* For Ascend950PR/Ascend950DT: support "NDHWC" and "NCDHW". \n71+* For Ascend 950PR&950DT products: support "NDHWC" and "NCDHW". \n
71 72 
72* @par Outputs:73* @par Outputs:
73* output: A mutable tensor with the same shape as "orig_input_shape" and same type as "grads".74* output: A mutable tensor with the same shape as "orig_input_shape" and same type as "grads".
Mpooling/avg_pool_grad/op_graph/avg_pool_grad_proto.h+3-3
@@ -27,16 +27,16 @@ namespace ge {
27* @par Attributes:27* @par Attributes:
28* @li ksize: A required tuple or list of ints,28* @li ksize: A required tuple or list of ints,
29* specifying the size of the window for each dimension of the input tensor.29* specifying the size of the window for each dimension of the input tensor.
30-* For Ascend 950PR/Ascend 950DT AI Processor: "ksize" length is 1, 2 or 4, must be greater than 0. \n30+* For Ascend 950PR&950DT products: "ksize" length is 1, 2 or 4, must be greater than 0. \n
31* @li strides: A required tuple or list of ints,31* @li strides: A required tuple or list of ints,
32* specifying the stride of the sliding window for each dimension of the input tensor.32* specifying the stride of the sliding window for each dimension of the input tensor.
33-* For Ascend 950PR/Ascend 950DT AI Processor: "strides" length is 1, 2 or 4, must be greater than 0. \n33+* For Ascend 950PR&950DT products: "strides" length is 1, 2 or 4, must be greater than 0. \n
34* @li padding: An optional string, specifying the type of the padding algorithm to use,34* @li padding: An optional string, specifying the type of the padding algorithm to use,
35* either "VALID", "SAME".35* either "VALID", "SAME".
36* With "SAME" means that the outputs will have the same spatial dimensions as its inputs.36* With "SAME" means that the outputs will have the same spatial dimensions as its inputs.
37* With "VALID" means no padding.37* With "VALID" means no padding.
38* @li data_format: An optional string. Defaults to "NHWC". \n38* @li data_format: An optional string. Defaults to "NHWC". \n
39-* For Ascend 950PR/Ascend 950DT AI Processor: support "NCHW" or "NHWC". \n39+* For Ascend 950PR&950DT products: support "NCHW" or "NHWC". \n
40 40 
41* @par Outputs:41* @par Outputs:
42* @li out_grad: A mutable tensor with the same shape as "orig_input_shape" and the same type as "input_grad". \n42* @li out_grad: A mutable tensor with the same shape as "orig_input_shape" and the same type as "input_grad". \n
Mpooling/avg_pool_v2_grad/op_graph/avg_pool_v2_grad_proto.h+3-3
@@ -27,10 +27,10 @@ namespace ge {
27* @par Attributes:27* @par Attributes:
28* @li ksize: A required tuple or list of ints,28* @li ksize: A required tuple or list of ints,
29* specifying the size of the window for each dimension of the input tensor.29* specifying the size of the window for each dimension of the input tensor.
30-* For Ascend 950PR/Ascend 950DT: "ksize" length is 4, must be greater than 0. \n30+* For Ascend 950PR&950DT products: "ksize" length is 4, must be greater than 0. \n
31* @li strides: A required tuple or list of ints,31* @li strides: A required tuple or list of ints,
32* specifying the stride of the sliding window for each dimension of the input tensor.32* specifying the stride of the sliding window for each dimension of the input tensor.
33-* For Ascend 950PR/Ascend 950DT: "strides" length is 4, must be greater than 0. \n33+* For Ascend 950PR&950DT products: "strides" length is 4, must be greater than 0. \n
34* @li padding_mode: An optional string, specifying the type of the padding algorithm to use,34* @li padding_mode: An optional string, specifying the type of the padding algorithm to use,
35* either "VALID", "SAME" or "CALCULATED". Default "CALCULATED".35* either "VALID", "SAME" or "CALCULATED". Default "CALCULATED".
36* With "SAME" means that the outputs will have the same spatial dimensions as its inputs.36* With "SAME" means that the outputs will have the same spatial dimensions as its inputs.
@@ -43,7 +43,7 @@ namespace ge {
43* and width. Default False.43* and width. Default False.
44* @li exclusive: An optional bool, whether to exclude padding points. Default is true.44* @li exclusive: An optional bool, whether to exclude padding points. Default is true.
45* @li data_format: An optional string. Defaults to "NCHW". \n45* @li data_format: An optional string. Defaults to "NCHW". \n
46-* For Ascend 950PR/Ascend 950DT: support "NCHW" or "NHWC". \n46+* For Ascend 950PR&950DT products: support "NCHW" or "NHWC". \n
47* @li divisor_override: An optional int, the default value is zero.47* @li divisor_override: An optional int, the default value is zero.
48* if specified, it will be used as divisor, otherwise size of the pooling region will be used.48* if specified, it will be used as divisor, otherwise size of the pooling region will be used.
49 49 
Mpooling/max_pool3_d/op_graph/max_pool3_d_proto.h+6-6
@@ -20,9 +20,9 @@ namespace ge {
20 20 
21* @par Inputs:21* @par Inputs:
22* One input:22* One input:
23-* x: A 5D tensor. Supported type:float16, float32. Additional support for bfloat16 in Ascend 950 AI Processor.23+* x: A 5D tensor. Supported type:float16, float32. Additional support for bfloat16 in Ascend 950PR&950DT products.
24* The double type is reserved but currently unsupported. Support format: NDHWC, NCDHW. Additional support for ND in24* The double type is reserved but currently unsupported. Support format: NDHWC, NCDHW. Additional support for ND in
25-Ascend 950 AI Processor.25+Ascend 950PR&950DT products.
26 26 
27* @par Attributes:27* @par Attributes:
28* @li ksize: A required list of int8, int16, int32, or int64 values,28* @li ksize: A required list of int8, int16, int32, or int64 values,
@@ -50,12 +50,12 @@ Ascend 950 AI Processor.
50* @li "ksize" is a list that has length 1, 3, or 5. The ksize of the H and W dimensions should be greater than 0.50* @li "ksize" is a list that has length 1, 3, or 5. The ksize of the H and W dimensions should be greater than 0.
51* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCDHW", ksize[0] = 1 and ksize[1] = 1.51* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCDHW", ksize[0] = 1 and ksize[1] = 1.
52* For "data_format" is "NDHWC", ksize[0] = 1 and ksize[4] = 1. \n52* For "data_format" is "NDHWC", ksize[0] = 1 and ksize[4] = 1. \n
53-* For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,53+* For Atlas training products, Atlas A2 products,
54-* Atlas A3 Training Series Product: The product of the ksize in D, H and W dimensions54+* Atlas A3 training products: The product of the ksize in D, H and W dimensions
55* should be less than or equal to 255. e.g. For "data_format" is "NCDHW", ksize[2] * ksize[3] * ksize[4] <= 255. \n55* should be less than or equal to 255. e.g. For "data_format" is "NCDHW", ksize[2] * ksize[3] * ksize[4] <= 255. \n
56* @li "strides" is a list that has length 1, 3, or 5. The stride of the N and C dimensions should be 1. \n56* @li "strides" is a list that has length 1, 3, or 5. The stride of the N and C dimensions should be 1. \n
57-* For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,57+* For Atlas training products, Atlas A2 products,
58-* Atlas A3 Training Series Product: The stride of the D, H and W dimensions should be greater than 0 and58+* Atlas A3 training products: The stride of the D, H and W dimensions should be greater than 0 and
59* smaller than 64. \n59* smaller than 64. \n
60* The stride of the D, H and W dimensions should be greater than 0.60* The stride of the D, H and W dimensions should be greater than 0.
61* @li "data_format" only support "NCDHW" and "NDHWC". \n61* @li "data_format" only support "NCDHW" and "NDHWC". \n
Mpooling/max_pool3d_grad_with_argmax/op_graph/max_pool_grad_with_argmax_v1_proto.h+1-1
@@ -22,7 +22,7 @@ namespace ge {
22* @li x: A tensor of type float16,float32, support format: [NC1HWC0, NCHW].22* @li x: A tensor of type float16,float32, support format: [NC1HWC0, NCHW].
23* @li grad: A tensor of type float16,float32, support format: [NC1HWC0, NCHW].23* @li grad: A tensor of type float16,float32, support format: [NC1HWC0, NCHW].
24* @li argmax: A tensor of type uint16,int32, support format: [NC1HWC0, NCHW]. \n24* @li argmax: A tensor of type uint16,int32, support format: [NC1HWC0, NCHW]. \n
25-* For Ascend950PR/Ascend950DT: The uint16 data type is not supported.25+* For Ascend 950PR&950DT products: The uint16 data type is not supported.
26 26 
27* @par Attributes:27* @par Attributes:
28* @li ksize: A required list of int8, int16, int32, or int64 values,28* @li ksize: A required list of int8, int16, int32, or int64 values,
Mpooling/max_pool_v3/op_graph/max_pool_proto.h+3-3
@@ -41,12 +41,12 @@ namespace ge {
41* @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.41* @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.
42* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.42* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.
43* For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n43* For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n
44-* For Non-Ascend950PR/Ascend950DT: The produce of the ksize in H and W dimensions44+* For products other than Ascend 950PR&950DT products: The produce of the ksize in H and W dimensions
45* should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n45* should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n
46* @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n46* @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n
47-* For Non-Ascend950PR/Ascend950DT: The stride of the H and W dimensions should be greater than 0 and47+* For products other than Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0 and
48* smaller than 64. \n48* smaller than 64. \n
49-* For Ascend950PR/Ascend950DT: The stride of the H and W dimensions should be greater than 0.49+* For Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0.
50* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output50* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output
51* shape at the H and W dimensions is calculated by below formula: \n51* shape at the H and W dimensions is calculated by below formula: \n
52* @code{.c}52* @code{.c}
Mpooling/max_pool_v3/op_graph/max_pool_v3_proto.h+6-6
@@ -56,14 +56,14 @@ namespace ge {
56* @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.56* @li "ksize" is a list that has length 4. The ksize of the H and W dimensions should be greater than 0.
57* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.57* The ksize of the N and C dimensions should be 1. e.g. For "data_format" is "NCHW", ksize[0] = 1 and ksize[1] = 1.
58* For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n58* For "data_format" is "NHWC", ksize[0] = 1 and ksize[3] = 1. \n
59-* For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,59+* For Atlas training products, Atlas A2 products,
60-* Atlas A3 Training Series Product: The produce of the ksize in H and W dimensions60+* Atlas A3 training products: The produce of the ksize in H and W dimensions
61* should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n61* should be less than or equal to 255. e.g. For "data_format" is "NCHW", ksize[2] * ksize[3] <= 255. \n
62* @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n62* @li "strides" is a list that has length 4. The stride of the N and C dimensions should be 1. \n
63-* For Atlas Training Series Product, Atlas A2 Training Series Product/Atlas 800I A2 Inference Product,63+* For Atlas training products, Atlas A2 products,
64-* Atlas A3 Training Series Product: The stride of the H and W dimensions should be greater than 0 and64+* Atlas A3 training products: The stride of the H and W dimensions should be greater than 0 and
65* smaller than 64. \n65* smaller than 64. \n
66-* For Ascend 950 AI Processor: The stride of the H and W dimensions should be greater than 0.66+* For Ascend 950PR&950DT products: The stride of the H and W dimensions should be greater than 0.
67* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output67* @li The ouput "y" shape at the N and C dimensions should be equal with input "x" shape at same dimensions. The output
68* shape at the H and W dimensions is calculated by below formula: \n68* shape at the H and W dimensions is calculated by below formula: \n
69* @code{.c}69* @code{.c}
@@ -108,4 +108,4 @@ REG_OP(MaxPoolV3)
108 .OP_END_FACTORY_REG(MaxPoolV3)108 .OP_END_FACTORY_REG(MaxPoolV3)
109 109 
110} // namespace ge110} // namespace ge
111-#endif111+#endif
Mquant/ascend_quant/op_graph/ascend_quant_proto.h+2-2
@@ -44,7 +44,7 @@ namespace ge {
44* Ceil: round to positive infinity(c language ceil). \n44* Ceil: round to positive infinity(c language ceil). \n
45* Trunc: round to zero(c language trunc). \n45* Trunc: round to zero(c language trunc). \n
46* Hybrid: only valid when output dtype is hifloat8. \n46* Hybrid: only valid when output dtype is hifloat8. \n
47-* The following constraints apply to products other than Ascend 950 AI Processor: \n47+* The following constraints apply to products other than Ascend 950PR&950DT products: \n
48* @li When format is FRACTAL_NZ, shape supports 4D ~ 8D.48* @li When format is FRACTAL_NZ, shape supports 4D ~ 8D.
49* @li When "x" is dynamic shape, shape [-2] is not supported.49* @li When "x" is dynamic shape, shape [-2] is not supported.
50* @li When "x" is dynamic shape, the data type of output "y" does not support int4.50* @li When "x" is dynamic shape, the data type of output "y" does not support int4.
@@ -63,4 +63,4 @@ REG_OP(AscendQuant)
63 .ATTR(dst_type, Int, DT_INT8)63 .ATTR(dst_type, Int, DT_INT8)
64 .OP_END_FACTORY_REG(AscendQuant)64 .OP_END_FACTORY_REG(AscendQuant)
65} // namespace ge65} // namespace ge
66-#endif66+#endif
Mquant/dequant_swiglu_quant/op_graph/dequant_swiglu_quant_proto.h+19-19
@@ -41,21 +41,21 @@ namespace ge {
41* 'false'(activate right) or 'true'(activate left), defalut is 'false'(activate right).41* 'false'(activate right) or 'true'(activate left), defalut is 'false'(activate right).
42* @li quant_mode: Type is string. The quant mode to use: 'static' or 'dynamic', defalut is 'static'.42* @li quant_mode: Type is string. The quant mode to use: 'static' or 'dynamic', defalut is 'static'.
43* @li dst_type: Type is Int32. Declare the output y dtype. Support 2:int8, 35:fp8_e5m2, 36:fp8_e4m3fn, 40:fp4x2_e2m1,43* @li dst_type: Type is Int32. Declare the output y dtype. Support 2:int8, 35:fp8_e5m2, 36:fp8_e4m3fn, 40:fp4x2_e2m1,
44-41:fp4x2_e1m2, Defaults to 2, only used for Ascend 950 AI Processors.44+41:fp4x2_e1m2, Defaults to 2, only used for Ascend 950PR&950DT products.
45* @li round_mode: Type is String. The round mode to use: 'rint', 'round, 'floor', 'ceil', 'trunc', defalut is 'rint',45* @li round_mode: Type is String. The round mode to use: 'rint', 'round, 'floor', 'ceil', 'trunc', defalut is 'rint',
46-only used for Ascend 950 AI Processors.46+only used for Ascend 950PR&950DT products.
47* @li activate_dim: Type is Int32. Describing the split dimension in Glu algorithm: value in [-xDim, xDim-1], defalut is47* @li activate_dim: Type is Int32. Describing the split dimension in Glu algorithm: value in [-xDim, xDim-1], defalut is
48--1, only used for Ascend 950 AI Processors.48+-1, only used for Ascend 950PR&950DT products.
49* @li swiglu_mode: Type is int. Optional parameter, default is 0. The SWIGLU computation mode to use:49* @li swiglu_mode: Type is int. Optional parameter, default is 0. The SWIGLU computation mode to use:
50* '0' (default) for standard SWIGLU, '1' for a variant using odd-even blocking, '2' for the same variant as '1' but50* '0' (default) for standard SWIGLU, '1' for a variant using odd-even blocking, '2' for the same variant as '1' but
51using contiguous front/back-half blocking (the same split as mode '0'). Modes '1' and '2' require clamp_limit,51using contiguous front/back-half blocking (the same split as mode '0'). Modes '1' and '2' require clamp_limit,
52-activation coefficient (glu_alpha), and bias (glu_bias). This attribute is not supported in Ascend 950 AI Processors.52+activation coefficient (glu_alpha), and bias (glu_bias). This attribute is not supported in Ascend 950PR&950DT products.
53* @li clamp_limit: Type is float. Optional parameter, default is 7.0. The threshold limit for SWIGLU input. This53* @li clamp_limit: Type is float. Optional parameter, default is 7.0. The threshold limit for SWIGLU input. This
54-attribute is not supported in Ascend 950 AI Processors.54+attribute is not supported in Ascend 950PR&950DT products.
55* @li glu_alpha: Type is float. Optional parameter, default is 1.702. The activation coefficient for the GLU activation55* @li glu_alpha: Type is float. Optional parameter, default is 1.702. The activation coefficient for the GLU activation
56-function. This attribute is not supported in Ascend 950 AI Processors.56+function. This attribute is not supported in Ascend 950PR&950DT products.
57* @li glu_bias: Type is float. Optional parameter, default is 1.0. The bias applied during SWIGLU linear computation.57* @li glu_bias: Type is float. Optional parameter, default is 1.0. The bias applied during SWIGLU linear computation.
58-This attribute is not supported in Ascend 950 AI Processors.58+This attribute is not supported in Ascend 950PR&950DT products.
59 59 
60* @attention Constraints:60* @attention Constraints:
61* @li When the type of x is int32, weight_scale must be input.61* @li When the type of x is int32, weight_scale must be input.
@@ -65,17 +65,17 @@ This attribute is not supported in Ascend 950 AI Processors.
65* @li When dst_type is hifloat8, round_mode supports 'round'.65* @li When dst_type is hifloat8, round_mode supports 'round'.
66* @li The shape of activate_dim corresponding to x must be divisible by 2.66* @li The shape of activate_dim corresponding to x must be divisible by 2.
67* @li When activate_dim is not the last dimension of x, group_index must be None.67* @li When activate_dim is not the last dimension of x, group_index must be None.
68-* @li The input quant_offset is not supported in Ascend 950 AI Processors only.68+* @li The input quant_offset is not supported in Ascend 950PR&950DT products only.
69-* @li The type of output y is fp8_e5m2, fp8_e4m3fn, fp4x2_e2m1 and fp4x2_e1m2 only supported in Ascend 950 AI69+* @li The type of output y is fp8_e5m2, fp8_e4m3fn, fp4x2_e2m1 and fp4x2_e1m2 only supported in Ascend 950PR&950DT
70-Processors.70+products.
71-* @li The attribute quant_mode is only supported 'dynamic' in Ascend 950 AI Processors.71+* @li The attribute quant_mode is only supported 'dynamic' in Ascend 950PR&950DT products.
72-* @li The attribute dst_type is only supported in Ascend 950 AI Processors.72+* @li The attribute dst_type is only supported in Ascend 950PR&950DT products.
73-* @li The attribute round_mode is only supported in Ascend 950 AI Processors.73+* @li The attribute round_mode is only supported in Ascend 950PR&950DT products.
74-* @li The attribute activate_dim is only supported in Ascend 950 AI Processors.74+* @li The attribute activate_dim is only supported in Ascend 950PR&950DT products.
75-* @li The attribute swiglu_mode is not supported in Ascend 950 AI Processors.75+* @li The attribute swiglu_mode is not supported in Ascend 950PR&950DT products.
76-* @li The attribute clamp_limit is not supported in Ascend 950 AI Processors.76+* @li The attribute clamp_limit is not supported in Ascend 950PR&950DT products.
77-* @li The attribute glu_alpha is not supported in Ascend 950 AI Processors.77+* @li The attribute glu_alpha is not supported in Ascend 950PR&950DT products.
78-* @li The attribute glu_bias is not supported in Ascend 950 AI Processors.78+* @li The attribute glu_bias is not supported in Ascend 950PR&950DT products.
79 79 
80* @par Restrictions:80* @par Restrictions:
81* Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.81* Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
@@ -102,4 +102,4 @@ REG_OP(DequantSwigluQuant)
102 .OP_END_FACTORY_REG(DequantSwigluQuant)102 .OP_END_FACTORY_REG(DequantSwigluQuant)
103} // namespace ge103} // namespace ge
104 104 
105-#endif // OPS_QUANT_DEQUANT_SWIGLU_QUANT_PROTO_H_105+#endif // OPS_QUANT_DEQUANT_SWIGLU_QUANT_PROTO_H_
Mquant/dynamic_quant/op_graph/dynamic_quant_proto.h+2-3
@@ -20,8 +20,7 @@ namespace ge {
20/**20/**
21 * @brief Dynamic Quant. Performs pre-token symmetric dynamic quantization on input tensors.21 * @brief Dynamic Quant. Performs pre-token symmetric dynamic quantization on input tensors.
22 * @par Inputs:22 * @par Inputs:
23- * @li x: A Tensor. Type is:DT_FLOAT16 or DT_BF16. For Atlas A2 Training Series Product/Atlas 800I A2 Inference23+ * @li x: A Tensor. Type is:DT_FLOAT16 or DT_BF16. For Atlas A2 products and Atlas A3 products.
24- * Product/A200I A2 Box Heterogeneous Component and Atlas A3 Training Series Product/Atlas A3 Inference Series Product.
25 * Whose shape must be greater than 1. The data format support ND.24 * Whose shape must be greater than 1. The data format support ND.
26 * @li smooth_scales: An optional Tensor.25 * @li smooth_scales: An optional Tensor.
27 * When group_index is null and smooth_scales is not null, shape is 1 Dims. Dim[0] is the last dimension of x.26 * When group_index is null and smooth_scales is not null, shape is 1 Dims. Dim[0] is the last dimension of x.
@@ -51,4 +50,4 @@ REG_OP(DynamicQuant)
51 .OP_END_FACTORY_REG(DynamicQuant)50 .OP_END_FACTORY_REG(DynamicQuant)
52} // namespace ge51} // namespace ge
53 52 
54-#endif // OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_53+#endif // OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_
Mquant/dynamic_quant_v2/op_graph/dynamic_quant_v2_proto.h+2-3
@@ -20,8 +20,7 @@ namespace ge {
20/**20/**
21 * @brief Dynamic Quant V2. Performs pre-token/per-tensor asymmetric dynamic quantization on input tensors.21 * @brief Dynamic Quant V2. Performs pre-token/per-tensor asymmetric dynamic quantization on input tensors.
22 * @par Inputs:22 * @par Inputs:
23- * @li x: A tensor. Type is:DT_FLOAT16 or DT_BF16. For Atlas A2 Training Series Product/Atlas 800I A2 Inference23+ * @li x: A tensor. Type is:DT_FLOAT16 or DT_BF16. For Atlas A2 products and Atlas A3 products.
24- * Product/A200I A2 Box Heterogeneous Component and Atlas A3 Training Series Product/Atlas A3 Inference Series Product.
25 * Whose shape must be greater than 1. The data format support ND.24 * Whose shape must be greater than 1. The data format support ND.
26 * @li smooth_scales: An optional tensor.25 * @li smooth_scales: An optional tensor.
27 * When group_index is null, shape is 1 Dims. Dim[0] is the last dimension of x.26 * When group_index is null, shape is 1 Dims. Dim[0] is the last dimension of x.
@@ -73,4 +72,4 @@ REG_OP(DynamicQuantV2)
73 .OP_END_FACTORY_REG(DynamicQuantV2)72 .OP_END_FACTORY_REG(DynamicQuantV2)
74} // namespace ge73} // namespace ge
75 74 
76-#endif // OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_75+#endif // OPS_BUILT_IN_OP_PROTO_INC_NN_QUANTIZE_H_
Mquant/trans_quant_param_v2/op_graph/trans_quant_param_v2_proto.h+6-6
@@ -37,12 +37,12 @@ namespace ge {
37* @attention Constraints:37* @attention Constraints:
38* @li The passed scale, y cannot be a null pointer.38* @li The passed scale, y cannot be a null pointer.
39 39 
40-* - Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or40+* - Atlas A2 products or
41-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product or Atlas Inference Series Product or41+* Atlas A3 products or Atlas inference products or
42-* Ascend 950 AI Processor: \n42+* Ascend 950PR&950DT products: \n
43* This operator supports use with the matmul operator, such as QuantBatchMatmulV3. \n43* This operator supports use with the matmul operator, such as QuantBatchMatmulV3. \n
44-* - Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component or44+* - Atlas A2 products or
45-* Atlas A3 Training Series Product/Atlas A3 Inference Series Product or Atlas Inference Series Product: \n45+* Atlas A3 products or Atlas inference products: \n
46* This operator does not supports use with the grouped matmul operator, such as GroupedMatmul. \n46* This operator does not supports use with the grouped matmul operator, such as GroupedMatmul. \n
47* @li When there is no offset, the y shape is consistent with the scale shape: \n47* @li When there is no offset, the y shape is consistent with the scale shape: \n
48* - If y is used as matmul input(e.g., QuantBatchMatmulV3), the shape support 1D (t,), with t equal to 1 or n, or 2D(1,48* - If y is used as matmul input(e.g., QuantBatchMatmulV3), the shape support 1D (t,), with t equal to 1 or n, or 2D(1,
@@ -64,4 +64,4 @@ REG_OP(TransQuantParamV2)
64 64 
65} // namespace ge65} // namespace ge
66 66 
67-#endif67+#endif
Mrnn/dynamic_rnnv2/op_graph/dynamic_rnnv2_proto.h+2-2
@@ -55,8 +55,8 @@ namespace ge {
55* the op. Default to "tanh". Only "tanh" is currently supported.55* the op. Default to "tanh". Only "tanh" is currently supported.
56* @li recurrent_activation:An string identifying the type of activation56* @li recurrent_activation:An string identifying the type of activation
57* function in the op. Default to "sigmoid". Only support "sigmoid" in57* function in the op. Default to "sigmoid". Only support "sigmoid" in
58-* Atlas A2 Training Series Product/Atlas 800I A2 Inference Product/A200I A2 Box Heterogeneous Component58+* Atlas A2 products
59-* and Atlas A3 Training Series Product/Atlas A3 Inference Series Product.59+* and Atlas A3 products.
60* Support "sigmoid" and "hard_sigmoid"60* Support "sigmoid" and "hard_sigmoid"
61* in other series produces. In general, set "hard_sigmoid" for TF Keras LSTM.61* in other series produces. In general, set "hard_sigmoid" for TF Keras LSTM.
62* @li forget_bias:An float identifying the forget bias in the op. Default to 0.62* @li forget_bias:An float identifying the forget bias in the op. Default to 0.