Merge "Supported FMA Intrinsics in Audio Resampler FIR processing module" am: a2fa90baca (865d6023) · Commits · e / os / android_frameworks_av

media/libaudioprocessing/Android.bp

+19 −0

Original line number	Diff line number	Diff line
		@@ -19,6 +19,25 @@ cc_defaults {
		// uncomment to disable NEON on architectures that actually do support NEON, for benchmarking
		// "-DUSE_NEON=false",
		],

		arch: {
		x86: {
		avx2: {
		cflags: [
		"-mavx2",
		"-mfma",
		],
		},
		},
		x86_64: {
		avx2: {
		cflags: [
		"-mavx2",
		"-mfma",
		],
		},
		},
		},
		}

		cc_library_shared {

media/libaudioprocessing/AudioResamplerFirOps.h

+8 −1

Original line number	Diff line number	Diff line
		@@ -36,13 +36,20 @@ namespace android {
		#include <arm_neon.h>
		#endif

		#if defined(__SSSE3__) // Should be supported in x86 ABI for both 32 & 64-bit.
		#if defined(__AVX2__) // Should be supported in x86 ABI for both 32 & 64-bit.
		#define USE_AVX2 (true) // Inference AVX2/FMA Intrinsics
		#define USE_SSE (true)
		#include <immintrin.h>
		#elif defined(__SSSE3__) // Should be supported in x86 ABI for both 32 & 64-bit.
		#define USE_SSE (true) // Inference SSE Intrinsics
		#define USE_AVX2 (false)
		#include <tmmintrin.h>
		#else
		#define USE_SSE (false)
		#define USE_AVX2(false)
		#endif


		template<typename T, typename U>
		struct is_same
		{

media/libaudioprocessing/AudioResamplerFirProcessSSE.h

+34 −11

Original line number	Diff line number	Diff line
		@@ -80,11 +80,16 @@ static inline void ProcessSSEIntrinsic(float* out,
		posCoef1 = _mm_sub_ps(posCoef1, posCoef);
		negCoef = _mm_sub_ps(negCoef, negCoef1);


		#if USE_AVX2
		posCoef = _mm_fmadd_ps(posCoef1, interp, posCoef);
		negCoef = _mm_fmadd_ps(negCoef, interp, negCoef1);
		#else
		posCoef1 = _mm_mul_ps(posCoef1, interp);
		negCoef = _mm_mul_ps(negCoef, interp);

		posCoef = _mm_add_ps(posCoef1, posCoef);
		negCoef = _mm_add_ps(negCoef, negCoef1);
		#endif //USE_AVX2
		}
		switch (CHANNELS) {
		case 1: {
		@@ -94,11 +99,17 @@ static inline void ProcessSSEIntrinsic(float* out,
		sN += 4;

		posSamp = _mm_shuffle_ps(posSamp, posSamp, 0x1B);

		#if USE_AVX2
		accL = _mm_fmadd_ps(posSamp, posCoef, accL);
		accL = _mm_fmadd_ps(negSamp, negCoef, accL);
		#else
		posSamp = _mm_mul_ps(posSamp, posCoef);
		negSamp = _mm_mul_ps(negSamp, negCoef);

		accL = _mm_add_ps(accL, posSamp);
		accL = _mm_add_ps(accL, negSamp);
		#endif

		} break;
		case 2: {
		__m128 posSamp0 = _mm_loadu_ps(sP);
		@@ -114,6 +125,12 @@ static inline void ProcessSSEIntrinsic(float* out,
		__m128 negSampL = _mm_shuffle_ps(negSamp0, negSamp1, 0x88);
		__m128 negSampR = _mm_shuffle_ps(negSamp0, negSamp1, 0xDD);

		#if USE_AVX2
		accL = _mm_fmadd_ps(posSampL, posCoef, accL);
		accR = _mm_fmadd_ps(posSampR, posCoef, accR);
		accL = _mm_fmadd_ps(negSampL, negCoef, accL);
		accR = _mm_fmadd_ps(negSampR, negCoef, accR);
		#else
		posSampL = _mm_mul_ps(posSampL, posCoef);
		posSampR = _mm_mul_ps(posSampR, posCoef);
		negSampL = _mm_mul_ps(negSampL, negCoef);
		@@ -123,6 +140,8 @@ static inline void ProcessSSEIntrinsic(float* out,
		accR = _mm_add_ps(accR, posSampR);
		accL = _mm_add_ps(accL, negSampL);
		accR = _mm_add_ps(accR, negSampR);
		#endif

		} break;
		}
		} while (count -= 4);
		@@ -144,9 +163,13 @@ static inline void ProcessSSEIntrinsic(float* out,
		outAccum = _mm_hadd_ps(accL, accR);
		outAccum = _mm_hadd_ps(outAccum, outAccum);
		}

		#if USE_AVX2
		outSamp = _mm_fmadd_ps(outAccum, vLR,outSamp);
		#else
		outAccum = _mm_mul_ps(outAccum, vLR);
		outSamp = _mm_add_ps(outSamp, outAccum);
		#endif

		_mm_storel_pi(reinterpret_cast<__m64*>(out), outSamp);
		}