add setdeqscalel4-pyasc-AtomGit

Mdocs/python-api/language/generated/asc.language.basic.set_deq_scale.md+19-0

	@@ -4,6 +4,8 @@
4		4
5	### asc.language.basic.set_deq_scale(scale: float, offset: int, sign_mode: bool) → None	5	### asc.language.basic.set_deq_scale(scale: float, offset: int, sign_mode: bool) → None
6		6
		7	+### asc.language.basic.set_deq_scale(vdeq: LocalTensor, vdeq_info: VdeqInfo) -> None:
		8	+
7	设置DEQSCALE寄存器的值。	9	设置DEQSCALE寄存器的值。
8		10
9	对应的Ascend C函数原型	11	对应的Ascend C函数原型
	@@ -12,6 +14,9 @@
12	__aicore__ inline void SetDeqScale(half scale)	14	__aicore__ inline void SetDeqScale(half scale)
13		15
14	__aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)	16	__aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)
		17	+
		18	+template <typename T>
		19	+__aicore__ inline void SetDeqScale(const LocalTensor<T>& vdeq, const VdeqInfo& vdeqInfo)
15	```	20	```
16		21
17	参数说明	22	参数说明
	@@ -20,6 +25,8 @@ __aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)
20	- scale(float)：scale量化参数，float类型。	25	- scale(float)：scale量化参数，float类型。
21	- offset：offset量化参数，int16_t类型，只有前9位有效。	26	- offset：offset量化参数，int16_t类型，只有前9位有效。
22	- sign_mode：bool类型，表示量化结果是否带符号。	27	- sign_mode：bool类型，表示量化结果是否带符号。
		28	+- vdeq: 输入量化tensor，大小为128Byte。类型为LocalTensor，支持的TPosition为VECIN/VECCALC/VECOUT。LocalTensor的起始地址需要32字节对齐。
		29	+- vdeq_info: 存储量化tensor信息的数据结构，结构体内包含量化tensor中的16组量化参数。
23		30
24	调用示例	31	调用示例
25		32
	@@ -28,6 +35,8 @@ __aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)
28	scale = 1.0	35	scale = 1.0
29	asc.set_deq_scale(scale)	36	asc.set_deq_scale(scale)
30	asc.cast(cast_dst_local, cast_dsrc_local, asc.RoundMode.CAST_NONE, src_size)	37	asc.cast(cast_dst_local, cast_dsrc_local, asc.RoundMode.CAST_NONE, src_size)
		38	+```
		39	+```python
31	# CastDeq	40	# CastDeq
32	scale = 1.0	41	scale = 1.0
33	offset = 0	42	offset = 0
	@@ -35,3 +44,13 @@ sign_mode = True
35	asc.set_deq_scale(scale, offset, sign_mode)	44	asc.set_deq_scale(scale, offset, sign_mode)
36	asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=False, half_block=False)	45	asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=False, half_block=False)
37	```	46	```
		47	+```python
		48	+# CastVdeq
		K K kong08081月9日调用示例建议隔开一些，或者分三个更清晰
		49	+vdeq_local = asc.LocalTensor(dtype=asc.uint64, pos=asc.TPosition.VECIN, addr=0, tile_size=16)
		50	+vdeq_scale = [1.0] * 16
		51	+vdeq_offset = [0] * 16
		52	+vdeq_sign_mode = [False] * 16
		53	+vdeq_info = asc.VdeqInfo(vdeq_scale, vdeq_offset, vdeq_sign_mode)
		54	+asc.set_deq_scale(vdeq_local, vdeq_info)
		55	+asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=True, half_block=False)
		56	+```

Minclude/ascir/API/Types.td+5-0

	@@ -438,4 +438,9 @@ def UnaryRepeatParams : APIType<"UnaryRepeatParams"> {
438	let mnemonic = "unary_repeat_params";	438	let mnemonic = "unary_repeat_params";
439	let apiName = "AscendC::UnaryRepeatParams";	439	let apiName = "AscendC::UnaryRepeatParams";
440	}	440	}
		441	+
		442	+def VdeqInfo : APIType<"VdeqInfo"> {
		443	+ let mnemonic = "vdeq_info";
		444	+ let apiName = "AscendC::VdeqInfo";
		445	+}
441	#endif // API_TYPES_TD	446	#endif // API_TYPES_TD

Minclude/ascir/Dialect/Asc/IR/Basic/OpVecVconv.td+6-0

	@@ -70,6 +70,12 @@ def AscendC_SetDeqScaleOp : APIOp<"set_deq_scale", "SetDeqScale", [AttrSizedOper
70	}];	70	}];
71	}	71	}
72		72
		73	+def AscendC_SetDeqScaleL4Op : VectorOp<"set_deq_scale_l4", "SetDeqScale"> {
		74	+ let description = "Set dequantization scale, offset and sign mode information";
		75	+ let arguments = (ins AscendC_LocalTensor:$vdeq, AscendC_VdeqInfo:$vdeqInfo
		76	+ );
		77	+}
		78	+
73	defm AddReluCast : BinaryCastL012Op<"add_relu_cast", "AddReluCast">;	79	defm AddReluCast : BinaryCastL012Op<"add_relu_cast", "AddReluCast">;
74	defm SubReluCast : BinaryCastL012Op<"sub_relu_cast", "SubReluCast">;	80	defm SubReluCast : BinaryCastL012Op<"sub_relu_cast", "SubReluCast">;
75		81

Minclude/ascir/Dialect/Asc/IR/Interfaces.td+2-0

	@@ -47,6 +47,8 @@ class AscendC_InterfaceMethods {
47	"getIsExhaustedSuspension", "bool">;	47	"getIsExhaustedSuspension", "bool">;
48	InterfaceMethod getIsFullSort = GetMethod<"isFullSort",	48	InterfaceMethod getIsFullSort = GetMethod<"isFullSort",
49	"getIsFullSort", "bool">;	49	"getIsFullSort", "bool">;
		50	+ InterfaceMethod getVdeq = GetMethod<"vdeq", "getVdeq">;
		51	+ InterfaceMethod getVdeqInfo = GetMethod<"vdeqInfo", "getVdeqInfo">;
50	}	52	}
51		53
52	class AscendC_OpInterface<string name, list<Interface> baseInterfaces = []>	54	class AscendC_OpInterface<string name, list<Interface> baseInterfaces = []>

Minclude/ascir/Target/Asc/Basic/VecVconv.h+1-0

	@@ -76,6 +76,7 @@ LogicalResult printOperation(CodeEmitter &emitter, ascendc::CastDeqL2Op op);
76		76
77	LogicalResult printOperation(CodeEmitter &emitter, ascendc::SetDeqScaleOp op);	77	LogicalResult printOperation(CodeEmitter &emitter, ascendc::SetDeqScaleOp op);
78		78
		79	+LogicalResult printOperation(CodeEmitter& emitter, ascendc::SetDeqScaleL4Op op);
79	} // namespace ascendc	80	} // namespace ascendc
80	} // namespace mlir	81	} // namespace mlir
81		82

Mlib/Target/AscendC/Basic/VecVconv.cpp+9-0

	@@ -86,3 +86,12 @@ LogicalResult mlir::ascendc::printOperation(CodeEmitter &emitter, ascendc::CastD
86	printUnaryL2Params(emitter, op);	86	printUnaryL2Params(emitter, op);
87	return success();	87	return success();
88	}	88	}
		89	+
		90	+LogicalResult mlir::ascendc::printOperation(CodeEmitter &emitter, ascendc::SetDeqScaleL4Op op) {
		91	+ auto &os = emitter.ostream();
		92	+ os << ascNamespace << "::" << op.getAPIName() << "("
		93	+ << emitter.getOrCreateName(op.getVdeq()) << ", "
		94	+ << emitter.getOrCreateName(op.getVdeqInfo()) << ")";
		95	+
		96	+ return success();
		97	+}

Mlib/Target/AscendC/Translation.cpp+1-1

	@@ -215,7 +215,7 @@ using PrintableOpTypes = std::tuple<
215	ascendc::RsqrtL2Op, ascendc::SqrtL2Op, ascendc::NegL2Op,	215	ascendc::RsqrtL2Op, ascendc::SqrtL2Op, ascendc::NegL2Op,
216	// VecVcon (Type conversion) operations	216	// VecVcon (Type conversion) operations
217	ascendc::CastL0Op, ascendc::CastL1Op, ascendc::CastL2Op, ascendc::CastDeqL0Op, ascendc::CastDeqL1Op,	217	ascendc::CastL0Op, ascendc::CastL1Op, ascendc::CastL2Op, ascendc::CastDeqL0Op, ascendc::CastDeqL1Op,
218	- ascendc::CastDeqL2Op,	218	+ ascendc::CastDeqL2Op, ascendc::SetDeqScaleL4Op,
219	// Vector gatherMask operations	219	// Vector gatherMask operations
220	ascendc::GatherMaskOp, ascendc::GetGatherMaskRemainCountOp,	220	ascendc::GatherMaskOp, ascendc::GetGatherMaskRemainCountOp,
221		221

Mpython/asc/language/__init__.py+1-0

	@@ -275,6 +275,7 @@ from .core.types import (
275	LoadData3DParamsV2Pro,	275	LoadData3DParamsV2Pro,
276	LoadDataRepeatParam,	276	LoadDataRepeatParam,
277	get_shape_size,	277	get_shape_size,
		278	+ VdeqInfo,
278	)	279	)
279	from .core.aipp_types import (	280	from .core.aipp_types import (
280	AippParams,	281	AippParams,

Mpython/asc/language/basic/utils.py+12-0

	@@ -3557,6 +3557,8 @@ def set_deq_scale_docstring():
3557		3557
3558	__aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)	3558	__aicore__ inline void SetDeqScale(float scale, int16_t offset, bool signMode)
3559		3559
		3560	+ template <typename T>
		3561	+ __aicore__ inline void SetDeqScale(const LocalTensor<T>& vdeq, const VdeqInfo& vdeqInfo)
3560	"""	3562	"""
3561		3563
3562	param_list = """	3564	param_list = """
	@@ -3566,6 +3568,8 @@ def set_deq_scale_docstring():
3566	- scale(float)：scale量化参数，float类型。	3568	- scale(float)：scale量化参数，float类型。
3567	- offset：offset量化参数，int16_t类型，只有前9位有效。	3569	- offset：offset量化参数，int16_t类型，只有前9位有效。
3568	- sign_mode：bool类型，表示量化结果是否带符号。	3570	- sign_mode：bool类型，表示量化结果是否带符号。
		3571	+ - vdeq: 输入量化tensor，大小为128Byte。类型为LocalTensor，支持的TPosition为VECIN/VECCALC/VECOUT。LocalTensor的起始地址需要32字节对齐。
		3572	+ - vdeqInfo: 存储量化tensor信息的数据结构，结构体内包含量化tensor中的16组量化参数。
3569	"""	3573	"""
3570		3574
3571	py_example = """	3575	py_example = """
	@@ -3583,6 +3587,14 @@ def set_deq_scale_docstring():
3583	sign_mode = True	3587	sign_mode = True
3584	asc.set_deq_scale(scale, offset, sign_mode)	3588	asc.set_deq_scale(scale, offset, sign_mode)
3585	asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=False, half_block=False)	3589	asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=False, half_block=False)
		3590	+ # CastVdeq
		3591	+ vdeq_local = asc.LocalTensor(dtype=asc.uint64, pos=asc.TPosition.VECIN, addr=0, tile_size=16)
		3592	+ vdeq_scale = [1.0] * 16
		3593	+ vdeq_offset = [0] * 16
		3594	+ vdeq_sign_mode = [False] * 16
		3595	+ vdeq_info = asc.VdeqInfo(vdeq_scale, vdeq_offset, vdeq_sign_mode)
		3596	+ asc.set_deq_scale(vdeq_local, vdeq_info)
		3597	+ asc.cast_deq(dst_local, src_local, count=src_size, is_vec_deq=True, half_block=False)
3586	"""	3598	"""
3587		3599
3588	return [func_introduction, cpp_signature, param_list, "", "", py_example]	3600	return [func_introduction, cpp_signature, param_list, "", "", py_example]

Mpython/asc/language/basic/vec_vconv.py+23-9

	@@ -6,7 +6,7 @@
6	# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	6	# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7	# See LICENSE in the root of the software repository for the full text of the License.	7	# See LICENSE in the root of the software repository for the full text of the License.
8		8
9	-from typing import List, overload, Optional	9	+from typing import List, overload
10		10
11	from ..._C import ir	11	from ..._C import ir
12	from ..core.dtype import KnownTypes, KnownTypes as KT	12	from ..core.dtype import KnownTypes, KnownTypes as KT
	@@ -14,7 +14,7 @@ from ..core.enums import RoundMode
14	from ..core.ir_value import RuntimeBool, RuntimeInt, RuntimeFloat, materialize_ir_value as _mat	14	from ..core.ir_value import RuntimeBool, RuntimeInt, RuntimeFloat, materialize_ir_value as _mat
15	from ..core.tensor import LocalTensor	15	from ..core.tensor import LocalTensor
16	from ..core.utils import require_jit, global_builder, DefaultValued, OverloadDispatcher	16	from ..core.utils import require_jit, global_builder, DefaultValued, OverloadDispatcher
17	-from ..core.types import BinaryRepeatParams, UnaryRepeatParams	17	+from ..core.types import BinaryRepeatParams, UnaryRepeatParams, VdeqInfo
18	from .utils import op_impl, set_binary_docstring, set_common_docstring	18	from .utils import op_impl, set_binary_docstring, set_common_docstring
19	from .vec_unary import op_impl as unary_op_impl	19	from .vec_unary import op_impl as unary_op_impl
20		20
	@@ -125,15 +125,29 @@ def set_deq_scale(scale: float, offset: int, sign_mode: bool) -> None:
125	...	125	...
126		126
127		127
		128	+@overload
		129	+def set_deq_scale(vdeq: LocalTensor, vdeq_info: VdeqInfo) -> None:
		130	+ ...
		131	+
		132	+
128	@require_jit	133	@require_jit
129	@set_common_docstring(api_name="set_deq_scale")	134	@set_common_docstring(api_name="set_deq_scale")
130	-def set_deq_scale(scale: RuntimeFloat, offset: Optional[RuntimeInt] = None,	135	+def set_deq_scale(args, *kwargs) -> None:
131	- sign_mode: Optional[RuntimeBool] = None) -> None:	136	+ builder = global_builder.get_ir_builder()
132	- if offset is None and sign_mode is None:	137	+ dispatcher = OverloadDispatcher("set_deq_scale")
133	- global_builder.get_ir_builder().create_asc_SetDeqScaleOp(_mat(scale, KnownTypes.half).to_ir())	138	+ @dispatcher.register(vdeq=LocalTensor, vdeq_info=VdeqInfo)
134	- else:	139	+ def _(vdeq: LocalTensor, vdeq_info: VdeqInfo):
135	- global_builder.get_ir_builder().create_asc_SetDeqScaleOp(_mat(scale, KnownTypes.float32).to_ir(), \	140	+ builder.create_asc_SetDeqScaleL4Op(vdeq.to_ir(), vdeq_info.to_ir())
136	- _mat(offset, KnownTypes.int16).to_ir(), _mat(sign_mode, KnownTypes.bit).to_ir())	141	+ @dispatcher.register(scale=RuntimeFloat)
		142	+ def _(scale: RuntimeFloat):
		143	+ builder.create_asc_SetDeqScaleOp(_mat(scale, KnownTypes.half).to_ir())
		144	+ @dispatcher.register(scale=RuntimeFloat, offset=RuntimeInt, sign_mode=RuntimeBool)
		145	+ def _(scale: RuntimeFloat, offset: RuntimeInt, sign_mode: RuntimeBool):
		146	+ builder.create_asc_SetDeqScaleOp(_mat(scale, KnownTypes.float32).to_ir(),
		147	+ _mat(offset, KnownTypes.int16).to_ir(),
		148	+ _mat(sign_mode, KnownTypes.bit).to_ir())
		149	+ dispatcher(args, *kwargs)
		150	+
137		151
138		152
139	@overload	153	@overload

Mpython/asc/language/core/__init__.py+3-1

	@@ -114,7 +114,8 @@ from .types import (
114	LoadData3DParamsV2,	114	LoadData3DParamsV2,
115	LoadData3DParamsV2Pro,	115	LoadData3DParamsV2Pro,
116	LoadDataRepeatParam,	116	LoadDataRepeatParam,
117	- get_shape_size	117	+ get_shape_size,
		118	+ VdeqInfo,
118	)	119	)
119	from .utils import ceildiv, static_assert	120	from .utils import ceildiv, static_assert
120		121
	@@ -236,6 +237,7 @@ __all__ = [
236	"LoadData3DParamsV2Pro",	237	"LoadData3DParamsV2Pro",
237	"LoadDataRepeatParam",	238	"LoadDataRepeatParam",
238	"get_shape_size",	239	"get_shape_size",
		240	+ "VdeqInfo",
239	# .core.utils	241	# .core.utils
240	"ceildiv",	242	"ceildiv",
241	"static_assert",	243	"static_assert",

Mpython/asc/language/core/types.py+64-0

	@@ -1915,3 +1915,67 @@ class FixpipeParamsV220(IRValue):
1915	def to_ir(self) -> IRHandle:	1915	def to_ir(self) -> IRHandle:
1916	return self.handle	1916	return self.handle
1917		1917
		1918	+
		1919	+class VdeqInfo(IRValue):
		1920	+
		1921	+ @overload
		1922	+ def __init__(
		1923	+ self,
		1924	+ scale: List[float],
		1925	+ offset: List[int],
		1926	+ sign_mode: List[bool],
		1927	+ ) -> None:
		1928	+
		1929	+ ...
		1930	+
		1931	+ @overload
		1932	+ def __init__(self, handle: IRHandle) -> None:
		1933	+ """This contructor should not be called by user"""
		1934	+ ...
		1935	+
		1936	+
		1937	+ @require_jit
		1938	+ def __init__(
		1939	+ self,
		1940	+ scale: Optional[List[float]] = None,
		1941	+ offset: Optional[List[int]] = None,
		1942	+ sign_mode: Optional[List[bool]] = None,
		1943	+ handle: Optional[IRHandle] = None,
		1944	+ ) -> None:
		1945	+ if handle is not None:
		1946	+ self.handle = handle
		1947	+ return
		1948	+
		1949	+ if scale is None or offset is None or sign_mode is None:
		1950	+ raise ValueError("VdeqInfo requires scale / offset / sign_mode")
		1951	+ if len(scale) != 16 or len(offset) != 16 or len(sign_mode) != 16:
		1952	+ raise ValueError("VdeqInfo expects exactly 16 elements per field")
		1953	+
		1954	+ builder = global_builder.get_ir_builder()
		1955	+
		1956	+ from .array import array
		1957	+
		1958	+ scale_array = array(KnownTypes.float32, scale)
		1959	+ offset_array = array(KnownTypes.int16, offset)
		1960	+ sign_mode_array = array(KnownTypes.bit, [1 if x else 0 for x in sign_mode])
		1961	+
		1962	+ self.handle = builder.create_asc_ConstructOp(
		1963	+ builder.get_asc_VdeqInfoType(),
		1964	+ [
		1965	+ scale_array.to_ir(),
		1966	+ offset_array.to_ir(),
		1967	+ sign_mode_array.to_ir(),
		1968	+ ],
		1969	+ builder.get_type_array_attr([
		1970	+ scale_array.to_ir().get_type(),
		1971	+ offset_array.to_ir().get_type(),
		1972	+ sign_mode_array.to_ir().get_type(),
		1973	+ ]),
		1974	+ )
		1975	+
		1976	+ @classmethod
		1977	+ def from_ir(cls, handle: IRHandle) -> "VdeqInfo":
		1978	+ return cls(handle=handle)
		1979	+
		1980	+ def to_ir(self) -> IRHandle:
		1981	+ return self.handle

Mpython/test/unit/language/basic/test_common_api.py+6-1

	@@ -576,7 +576,12 @@ def test_set_deq_scale(mock_launcher_run):
576	def kernel_set_deq_scale() -> None:	576	def kernel_set_deq_scale() -> None:
577	asc.set_deq_scale(1.0)	577	asc.set_deq_scale(1.0)
578	asc.set_deq_scale(1.0, 5, False)	578	asc.set_deq_scale(1.0, 5, False)
579	-	579	+ vdeq_local = asc.LocalTensor(dtype=asc.uint64, pos=asc.TPosition.VECIN, addr=0, tile_size=16)
		580	+ vdeq_scale = [1.0] * 16
		581	+ vdeq_offset = [5] * 16
		582	+ vdeq_sign_mode = [False] * 16
		583	+ vdeq_info = asc.VdeqInfo(vdeq_scale, vdeq_offset, vdeq_sign_mode)
		584	+ asc.set_deq_scale(vdeq_local, vdeq_info)
580	kernel_set_deq_scale[1]()	585	kernel_set_deq_scale[1]()
581	assert mock_launcher_run.call_count == 1	586	assert mock_launcher_run.call_count == 1
582		587

Mtest/Target/AscendC/basic/vec_vconv.mlir+9-0

	@@ -70,6 +70,15 @@ func.func @emit_set_deq_scale(%arg0: f16, %arg1: f32, %arg2: i16) {
70	return	70	return
71	}	71	}
72		72
		73	+// CHECK-LABEL:void emit_set_deq_scale_l4(AscendC::LocalTensor<float> v1, AscendC::VdeqInfo v2) {
		74	+// CHECK-NEXT: AscendC::SetDeqScale(v1, v2);
		75	+// CHECK-NEXT: return;
		76	+// CHECK-NEXT: }
		77	+func.func @emit_set_deq_scale_l4(%vdeq: !ascendc.local_tensor<32xf32>, %vdeq_info: !ascendc.vdeq_info) {
		78	+ ascendc.set_deq_scale_l4 %vdeq, %vdeq_info : !ascendc.local_tensor<32xf32>, !ascendc.vdeq_info
		79	+ return
		80	+}
		81	+
73	// CHECK-LABEL:void emit_cast_deq(AscendC::LocalTensor<float> v1, AscendC::LocalTensor<float> v2, uint8_t v3, AscendC::UnaryRepeatParams v4, uint64_t v5, uint64_t v6, int32_t v7) {	82	// CHECK-LABEL:void emit_cast_deq(AscendC::LocalTensor<float> v1, AscendC::LocalTensor<float> v2, uint8_t v3, AscendC::UnaryRepeatParams v4, uint64_t v5, uint64_t v6, int32_t v7) {
74	// CHECK-NEXT: AscendC::CastDeq<float, float, 0, 0>(v1, v2, v7);	83	// CHECK-NEXT: AscendC::CastDeq<float, float, 0, 0>(v1, v2, v7);
75	// CHECK-NEXT: AscendC::CastDeq<float, float, 1, 0, 0>(v1, v2, v5, v3, v4);	84	// CHECK-NEXT: AscendC::CastDeq<float, float, 1, 0, 0>(v1, v2, v5, v3, v4);