Add BoxMullerFloat-ops-math-AtomGit

已合并

Add BoxMullerFloat #3071

梅国晗954517创建于 6月1日从gh_M/ops-math_0601： new_boxmullerfloat 合入到 cann/ops-math： master

Add BoxMullerFloat #3071

已合并

梅国晗954517创建于 6月1日

从gh_M/ops-math_0601 new_boxmullerfloat合入到cann/ops-math master

共 11 个文件变更+23-12

Mrandom/random_common/op_kernel/arch35/random_kernel_base.h+12-2

	@@ -311,8 +311,8 @@ __aicore__ inline void BoxMullerMulSIMD(
311	}	311	}
312	}	312	}
313		313
314	-// Box-Muller transform	314	+// Box-Muller transform（Use BoxMullerFloatSafe for eps-protected version)
315	-__simt_callee__ __aicore__ inline void BoxMullerFloat(float u1, const float u2, float* z0, float* z1)	315	+__simt_callee__ __aicore__ inline void BoxMullerFloatSafe(float u1, const float u2, float* z0, float* z1)
316	{	316	{
317	const float eps = 1.0e-7f;	317	const float eps = 1.0e-7f;
318	if (u1 < eps) {	318	if (u1 < eps) {
	@@ -325,6 +325,16 @@ __simt_callee__ __aicore__ inline void BoxMullerFloat(float u1, const float u2,
325	z1 = r;	325	z1 = r;
326	}	326	}
327		327
		328	+// Box-Muller transform (Use BoxMullerFloat for torch-aligned, no eps clamping. ）
		329	+__simt_callee__ __aicore__ inline void BoxMullerFloat(float u1, const float u2, float* z0, float* z1)
		SS S sunday6月3日将原 `BoxMullerFloat` 重命名为 `BoxMullerFloatSafe` 后，新增了同名但无 eps 保护的新 `BoxMullerFloat`。`stateless_normal.h`（不在本 PR 变更范围内）通过 `using namespace RandomKernelBase` 调用 `BoxMullerFloat(u1, u2, &z0, &z1)`，此前调用的是带 eps 保护的版本，此 PR 后隐式切换为无 eps 保护的对齐 torch 版本，行为发生变化。需要确认：1) `stateless_normal` 是否有意使用对齐 torch 的无 eps 版本（即当前隐式切换后的行为），若是请在该文件中显式标注或加注释说明；2) 若需保持原有安全行为，则应将该调用改为 `BoxMullerFloatSafe`。 `truncated_normal_v2` 已显式改为 `BoxMullerFloatSafe`（截断采样场景需要 eps 保护避免极端值浪费迭代），逻辑正确。但 `stateless_normal` 未做对应处理，存在遗漏风险。梅国晗954517 6月3日评论： 6月3日评论： S sunday6月3日新增的 `BoxMullerFloat`（无 eps 保护，对齐 torch）与 `BoxMullerFloatSafe`（有 eps 保护）共存，但缺少注释说明两者差异和适用场景。建议在函数上方加一行注释，如 `// Box-Muller transform (torch-aligned, no eps clamping. Use BoxMullerFloatSafe for eps-protected version)`，便于后续维护者理解。梅国晗954517 6月3日评论： 6月3日评论：
		330	+{
		331	+ float v = static_cast<float>(DOUBLE_MULTIPLE * PI * u2);
		332	+ float r = sqrtf(-DOUBLE_MULTIPLE * logf(u1));
		333	+ sincosf(v, z0, z1);
		334	+ z0 = r;
		335	+ z1 = r;
		336	+}
		337	+
328	template <uint16_t COPY_SIZE>	338	template <uint16_t COPY_SIZE>
329	__simt_callee__ __aicore__ inline void CopyArray(uint32_t* dst, const uint32_t* src)	339	__simt_callee__ __aicore__ inline void CopyArray(uint32_t* dst, const uint32_t* src)
330	{	340	{

Mrandom/stateless_normal/op_kernel/arch35/stateless_normal.h+1-0

	@@ -49,6 +49,7 @@ struct NormalTransform {
49	float u1 = results[pairBase] * RAND_2POW32_INV + RAND_2POW32_INV_HALF;	49	float u1 = results[pairBase] * RAND_2POW32_INV + RAND_2POW32_INV_HALF;
50	float u2 = results[pairBase + 1] * RAND_2POW32_INV + RAND_2POW32_INV_HALF;	50	float u2 = results[pairBase + 1] * RAND_2POW32_INV + RAND_2POW32_INV_HALF;
51	float z0, z1;	51	float z0, z1;
		52	+ // 使用对齐torch版本的BoxMullerFloat，无需eps保护
52	BoxMullerFloat(u1, u2, &z0, &z1);	53	BoxMullerFloat(u1, u2, &z0, &z1);
53	float z = (iStep % 2 == 0) ? z0 : z1;	54	float z = (iStep % 2 == 0) ? z0 : z1;
54	if constexpr (IsSameType<T, float>::value) {	55	if constexpr (IsSameType<T, float>::value) {

Mrandom/stateless_random_normal_v2/docs/aclnnNormalFloatFloat.md+1-1

	@@ -302,7 +302,7 @@ int main() {
302	float meanValue = 1.2f;	302	float meanValue = 1.2f;
303	float stdValue = 2.5f;	303	float stdValue = 2.5f;
304	int64_t seed = 1;	304	int64_t seed = 1;
305	- int64_t offset = 2;	305	+ int64_t offset = 4;
		S S sunday6月3日 4 个文档示例中 `offset = 4`，4 个 examples 代码中 `offset = 12`，两者值不同。虽然都满足"4 的倍数"约束，但文档和示例代码使用不同的 offset 值会让读者困惑。建议统一为一个值（如都用 4 或都用 12），或在 examples 中加注释说明为什么选 12 而不是 4。梅国晗954517 6月3日评论： 6月3日评论：
306		306
307	// 创建out aclTensor	307	// 创建out aclTensor
308	ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);	308	ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);

Mrandom/stateless_random_normal_v2/docs/aclnnNormalFloatTensor.md+1-1

	@@ -308,7 +308,7 @@ int main() {
308	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	308	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
309	float meanValue = 1.2f;	309	float meanValue = 1.2f;
310	int64_t seed = 1;	310	int64_t seed = 1;
311	- int64_t offset = 1;	311	+ int64_t offset = 4;
312	// 创建self aclTensor	312	// 创建self aclTensor
313	ret = CreateAclTensor(stdHostData, stdShape, &stdDeviceAddr, aclDataType::ACL_FLOAT, &std);	313	ret = CreateAclTensor(stdHostData, stdShape, &stdDeviceAddr, aclDataType::ACL_FLOAT, &std);
314	CHECK_RET(ret == ACL_SUCCESS, return ret);	314	CHECK_RET(ret == ACL_SUCCESS, return ret);

Mrandom/stateless_random_normal_v2/docs/aclnnNormalTensorFloat.md+1-1

	@@ -311,7 +311,7 @@ int main() {
311	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	311	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
312	float stdValue = 1.2f;	312	float stdValue = 1.2f;
313	int64_t seed = 1;	313	int64_t seed = 1;
314	- int64_t offset = 1;	314	+ int64_t offset = 4;
315	// 创建self aclTensor	315	// 创建self aclTensor
316	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);	316	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);
317	CHECK_RET(ret == ACL_SUCCESS, return ret);	317	CHECK_RET(ret == ACL_SUCCESS, return ret);

Mrandom/stateless_random_normal_v2/docs/aclnnNormalTensorTensor.md+1-1

	@@ -320,7 +320,7 @@ int main() {
320	std::vector<float> stdHostData = {0.5, 0.6, 0.4, 0.5};	320	std::vector<float> stdHostData = {0.5, 0.6, 0.4, 0.5};
321	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	321	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
322	int64_t seed = 1;	322	int64_t seed = 1;
323	- int64_t offset = 1;	323	+ int64_t offset = 4;
324		324
325	// 创建mean aclTensor	325	// 创建mean aclTensor
326	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);	326	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);

Mrandom/stateless_random_normal_v2/examples/test_aclnn_normal_float_float.cpp+1-1

	@@ -94,7 +94,7 @@ int main()
94	float meanValue = 1.2f;	94	float meanValue = 1.2f;
95	float stdValue = 2.5f;	95	float stdValue = 2.5f;
96	int64_t seed = 1;	96	int64_t seed = 1;
97	- int64_t offset = 2;	97	+ int64_t offset = 4;
98		98
99	// 创建out aclTensor	99	// 创建out aclTensor
100	ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);	100	ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);

Mrandom/stateless_random_normal_v2/examples/test_aclnn_normal_float_tensor.cpp+1-1

	@@ -97,7 +97,7 @@ int main()
97	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	97	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
98	float meanValue = 1.2f;	98	float meanValue = 1.2f;
99	int64_t seed = 1;	99	int64_t seed = 1;
100	- int64_t offset = 1;	100	+ int64_t offset = 4;
101	// 创建self aclTensor	101	// 创建self aclTensor
102	ret = CreateAclTensor(stdHostData, stdShape, &stdDeviceAddr, aclDataType::ACL_FLOAT, &std);	102	ret = CreateAclTensor(stdHostData, stdShape, &stdDeviceAddr, aclDataType::ACL_FLOAT, &std);
103	CHECK_RET(ret == ACL_SUCCESS, return ret);	103	CHECK_RET(ret == ACL_SUCCESS, return ret);

Mrandom/stateless_random_normal_v2/examples/test_aclnn_normal_tensor_float.cpp+1-1

	@@ -97,7 +97,7 @@ int main()
97	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	97	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
98	float stdValue = 1.2f;	98	float stdValue = 1.2f;
99	int64_t seed = 1;	99	int64_t seed = 1;
100	- int64_t offset = 1;	100	+ int64_t offset = 4;
101	// 创建self aclTensor	101	// 创建self aclTensor
102	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);	102	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);
103	CHECK_RET(ret == ACL_SUCCESS, return ret);	103	CHECK_RET(ret == ACL_SUCCESS, return ret);

Mrandom/stateless_random_normal_v2/examples/test_aclnn_normal_tensor_tensor.cpp+1-1

	@@ -100,7 +100,7 @@ int main()
100	std::vector<float> stdHostData = {0.5, 0.6, 0.4, 0.5};	100	std::vector<float> stdHostData = {0.5, 0.6, 0.4, 0.5};
101	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};	101	std::vector<float> outHostData = {0.0, 0.0, 0.0, 0.0};
102	int64_t seed = 1;	102	int64_t seed = 1;
103	- int64_t offset = 1;	103	+ int64_t offset = 4;
104		104
105	// 创建mean aclTensor	105	// 创建mean aclTensor
106	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);	106	ret = CreateAclTensor(meanHostData, meanShape, &meanDeviceAddr, aclDataType::ACL_FLOAT, &mean);

Mrandom/truncated_normal_v2/op_kernel/arch35/truncated_normal_v2_simt.h+2-2

	@@ -73,7 +73,7 @@ __simt_callee__ __aicore__ inline void GenSamples(float* results, const uint32_t
73	// the desired number of elements that fall within the pre-defined cutoff	73	// the desired number of elements that fall within the pre-defined cutoff
74	// threshold.	74	// threshold.
75	float f[2];	75	float f[2];
76	- RandomKernelBase::BoxMullerFloat(Uint32ToFloat(counterRst[0]), Uint32ToFloat(counterRst[1]), &f[0], &f[1]);	76	+ RandomKernelBase::BoxMullerFloatSafe(Uint32ToFloat(counterRst[0]), Uint32ToFloat(counterRst[1]), &f[0], &f[1]);
77	FilterSample(results, index, f[0]);	77	FilterSample(results, index, f[0]);
78	if (index >= static_cast<int>(GROUP_SIZE)) {	78	if (index >= static_cast<int>(GROUP_SIZE)) {
79	return;	79	return;
	@@ -83,7 +83,7 @@ __simt_callee__ __aicore__ inline void GenSamples(float* results, const uint32_t
83	return;	83	return;
84	}	84	}
85		85
86	- RandomKernelBase::BoxMullerFloat(Uint32ToFloat(counterRst[2]), Uint32ToFloat(counterRst[3]), &f[0], &f[1]);	86	+ RandomKernelBase::BoxMullerFloatSafe(Uint32ToFloat(counterRst[2]), Uint32ToFloat(counterRst[3]), &f[0], &f[1]);
87	FilterSample(results, index, f[0]);	87	FilterSample(results, index, f[0]);
88	if (index >= static_cast<int>(GROUP_SIZE)) {	88	if (index >= static_cast<int>(GROUP_SIZE)) {
89	return;	89	return;