CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutSign UpSign In
hrydgard

CoCalc provides the best real-time collaborative environment for Jupyter Notebooks, LaTeX documents, and SageMath, scalable from individual users to large groups and classes!

GitHub Repository: hrydgard/ppsspp
 Path: blob/master/Common/Math/CrossSIMD.h
Views: 1401
1
// CrossSIMD

2
//

3
// Compatibility wrappers for SIMD dialects.

4
//

5
// In the long run, might do a more general single-source-SIMD wrapper here consisting

6
// of defines that translate to either NEON or SSE. It would be possible to write quite a lot of

7
// our various color conversion functions and so on in a pretty generic manner.

8


9
#include "ppsspp_config.h"

10


11
#include "stdint.h"

12


13
#ifdef __clang__

14
// Weird how you can't just use #pragma in a macro.

15
#define DO_NOT_VECTORIZE_LOOP _Pragma("clang loop vectorize(disable)")

16
#else

17
#define DO_NOT_VECTORIZE_LOOP

18
#endif

19


20
#if PPSSPP_ARCH(SSE2)

21
#include <emmintrin.h>

22
#endif

23


24
#if PPSSPP_ARCH(ARM_NEON)

25
#if defined(_MSC_VER) && PPSSPP_ARCH(ARM64)

26
#include <arm64_neon.h>

27
#else

28
#include <arm_neon.h>

29
#endif

30
#endif

31


32
// Basic types

33


34
#if PPSSPP_ARCH(ARM64_NEON)

35


36
// No special ones here.

37


38
#elif PPSSPP_ARCH(ARM_NEON)

39


40
// Compatibility wrappers making ARM64 NEON code run on ARM32

41
// With optimization on, these should compile down to the optimal code.

42


43
inline float32x4_t vmulq_laneq_f32(float32x4_t a, float32x4_t b, int lane) {

44
	switch (lane & 3) {

45
	case 0: return vmulq_lane_f32(a, vget_low_f32(b), 0);

46
	case 1: return vmulq_lane_f32(a, vget_low_f32(b), 1);

47
	case 2: return vmulq_lane_f32(a, vget_high_f32(b), 0);

48
	default: return vmulq_lane_f32(a, vget_high_f32(b), 1);

49
	}

50
}

51


52
inline float32x4_t vmlaq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t c, int lane) {

53
	switch (lane & 3) {

54
	case 0: return vmlaq_lane_f32(a, b, vget_low_f32(c), 0);

55
	case 1: return vmlaq_lane_f32(a, b, vget_low_f32(c), 1);

56
	case 2: return vmlaq_lane_f32(a, b, vget_high_f32(c), 0);

57
	default: return vmlaq_lane_f32(a, b, vget_high_f32(c), 1);

58
	}

59
}

60


61
inline uint32x4_t vcgezq_f32(float32x4_t v) {

62
	return vcgeq_f32(v, vdupq_n_f32(0.0f));

63
}

64


65
#endif

66


67