浮点运算指令-asc-devkit-AtomGit

已合并

浮点运算指令 #1774

zhangshunwei2创建于 4月23日从zhangshunwei2/asc-devkit： floatcompute 合入到 cann/asc-devkit： master

浮点运算指令 #1774

已合并

zhangshunwei2创建于 4月23日

从zhangshunwei2/asc-devkit floatcompute合入到cann/asc-devkit master

共 3 个文件变更+104-0

Minclude/simt_api/asc_bf16.h+42-0

	@@ -474,6 +474,48 @@ static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __hgtux2_mask(bfloat16x2_t x,
474		474
475	static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __isnanx2(bfloat16x2_t x);	475	static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __isnanx2(bfloat16x2_t x);
476		476
		477	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __haddx2(const bfloat16x2_t x, const bfloat16x2_t y);
		478	+
		479	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hsubx2(const bfloat16x2_t x, const bfloat16x2_t y);
		480	+
		481	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hmulx2(const bfloat16x2_t x, const bfloat16x2_t y);
		482	+
		483	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hdivx2(const bfloat16x2_t x, const bfloat16x2_t y);
		484	+
		485	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __habsx2(const bfloat16x2_t x);
		486	+
		487	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hfmax2(const bfloat16x2_t x, const bfloat16x2_t y, const bfloat16x2_t z);
		488	+
		489	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hnegx2(const bfloat16x2_t x);
		490	+
		491	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hadd(const bfloat16_t x, const bfloat16_t y);
		492	+
		493	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hsub(const bfloat16_t x, const bfloat16_t y);
		494	+
		495	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hmul(const bfloat16_t x, const bfloat16_t y);
		496	+
		497	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hdiv(const bfloat16_t x, const bfloat16_t y);
		498	+
		499	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hneg(const bfloat16_t x);
		500	+
		501	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hmax_nan(const bfloat16_t x, const bfloat16_t y);
		502	+
		503	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hmin_nan(const bfloat16_t x, const bfloat16_t y);
		504	+
		505	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16_t __hfma_relu(const bfloat16_t x, const bfloat16_t y, const bfloat16_t z);
		506	+
		507	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hmaxx2(const bfloat16x2_t x, const bfloat16x2_t y);
		508	+
		509	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hminx2(const bfloat16x2_t x, const bfloat16x2_t y);
		510	+
		511	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hmaxx2_nan(const bfloat16x2_t x, const bfloat16x2_t y);
		512	+
		513	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hminx2_nan(const bfloat16x2_t x, const bfloat16x2_t y);
		514	+
		515	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hfmax2_relu(const bfloat16x2_t x, const bfloat16x2_t y, const bfloat16x2_t z);
		516	+
		517	+static __SIMT_DEVICE_FUNCTIONS_DECL__ bfloat16x2_t __hcmadd(const bfloat16x2_t x, const bfloat16x2_t y, const bfloat16x2_t z);
		518	+
477	#include "impl/simt_api/asc_bf16_impl.h"	519	#include "impl/simt_api/asc_bf16_impl.h"
478		520
479	#endif	521	#endif

Minclude/simt_api/asc_fp16.h+42-0

	@@ -486,6 +486,48 @@ static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __hgtux2_mask(half2 x, half2
486		486
487	static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __isnanx2(half2 x);	487	static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __isnanx2(half2 x);
488		488
		489	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hmaxx2(const half2 x, const half2 y);
		490	+
		491	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hminx2(const half2 x, const half2 y);
		492	+
		493	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __haddx2(const half2 x, const half2 y);
		494	+
		495	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hsubx2(const half2 x, const half2 y);
		496	+
		497	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hmulx2(const half2 x, const half2 y);
		498	+
		499	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hdivx2(const half2 x, const half2 y);
		500	+
		501	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __habsx2(const half2 x);
		502	+
		503	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hfmax2(const half2 x, const half2 y, const half2 z);
		504	+
		505	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hnegx2(const half2 x);
		506	+
		507	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hadd(const half x, const half y);
		508	+
		509	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hsub(const half x, const half y);
		510	+
		511	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hmul(const half x, const half y);
		512	+
		513	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hdiv(const half x, const half y);
		514	+
		515	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hneg(const half x);
		516	+
		517	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hmax_nan(const half x, const half y);
		518	+
		519	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hmin_nan(const half x, const half y);
		520	+
		521	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half __hfma_relu(const half x, const half y, const half z);
		522	+
		523	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hmaxx2_nan(const half2 x, const half2 y);
		524	+
		525	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hminx2_nan(const half2 x, const half2 y);
		526	+
		527	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hfmax2_relu(const half2 x, const half2 y, const half2 z);
		528	+
		529	+static __SIMT_DEVICE_FUNCTIONS_DECL__ half2 __hcmadd(const half2 x, const half2 y, const half2 z);
		530	+
489	#include "impl/simt_api/asc_fp16_impl.h"	531	#include "impl/simt_api/asc_fp16_impl.h"
490		532
491	#endif	533	#endif

Minclude/simt_api/device_functions.h+20-0

	@@ -442,6 +442,26 @@ static __SIMT_DEVICE_FUNCTIONS_DECL__ int __popc(unsigned long x);
442		442
443	static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __byte_perm(unsigned int x, unsigned int y, unsigned int s);	443	static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __byte_perm(unsigned int x, unsigned int y, unsigned int s);
444		444
		445	+static __SIMT_DEVICE_FUNCTIONS_DECL__ float __saturatef(float x);
		446	+
		447	+static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __sad(int x, int y, unsigned int z);
		448	+
		449	+static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __usad(unsigned int x, unsigned int y, unsigned int z);
		450	+
		451	+static __SIMT_DEVICE_FUNCTIONS_DECL__ int __mul24(int x, int y);
		452	+
		453	+static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __umul24(unsigned int x, unsigned int y);
		454	+
		455	+static __SIMT_DEVICE_FUNCTIONS_DECL__ float __fdividef(float x, float y);
		456	+
		457	+static __SIMT_DEVICE_FUNCTIONS_DECL__ int __hadd(int x, int y);
		458	+
		459	+static __SIMT_DEVICE_FUNCTIONS_DECL__ int __rhadd(int x, int y);
		460	+
		461	+static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __uhadd(unsigned int x, unsigned int y);
		462	+
		463	+static __SIMT_DEVICE_FUNCTIONS_DECL__ unsigned int __urhadd(unsigned int x, unsigned int y);
		464	+
445	#include "impl/simt_api/device_functions_impl.h"	465	#include "impl/simt_api/device_functions_impl.h"
446		466
447	#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_DEVICE_FUNCTIONS_H__)	467	#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_DEVICE_FUNCTIONS_H__)