Home
last modified time | relevance | path

Searched refs:_mm_loadu_si128 (Results 1 – 25 of 37) sorted by relevance

12

/external/webp/src/dsp/
Denc_sse2.c73 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); in CollectHistogramSSE2()
74 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); in CollectHistogramSSE2()
503 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]); in SSE_Nx4SSE2()
504 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]); in SSE_Nx4SSE2()
505 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]); in SSE_Nx4SSE2()
506 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]); in SSE_Nx4SSE2()
507 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]); in SSE_Nx4SSE2()
508 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]); in SSE_Nx4SSE2()
509 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]); in SSE_Nx4SSE2()
510 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]); in SSE_Nx4SSE2()
[all …]
Dupsampling_sse2.c67 const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \
68 const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \
69 const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \
70 const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \
Ddec_sse2.c601 __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]); in SimpleVFilter16SSE2()
602 __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]); in SimpleVFilter16SSE2()
603 __m128i q0 = _mm_loadu_si128((__m128i*)&p[0]); in SimpleVFilter16SSE2()
604 __m128i q1 = _mm_loadu_si128((__m128i*)&p[stride]); in SimpleVFilter16SSE2()
655 e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]); \
656 e2 = _mm_loadu_si128((__m128i*)&(p)[1 * stride]); \
657 e3 = _mm_loadu_si128((__m128i*)&(p)[2 * stride]); \
658 e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]); \
/external/chromium_org/third_party/libwebp/dsp/
Denc_sse2.c73 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); in CollectHistogramSSE2()
74 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); in CollectHistogramSSE2()
503 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]); in SSE_Nx4SSE2()
504 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]); in SSE_Nx4SSE2()
505 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]); in SSE_Nx4SSE2()
506 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]); in SSE_Nx4SSE2()
507 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]); in SSE_Nx4SSE2()
508 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]); in SSE_Nx4SSE2()
509 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]); in SSE_Nx4SSE2()
510 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]); in SSE_Nx4SSE2()
[all …]
Dupsampling_sse2.c67 const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \
68 const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \
69 const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \
70 const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \
Ddec_sse2.c601 __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]); in SimpleVFilter16SSE2()
602 __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]); in SimpleVFilter16SSE2()
603 __m128i q0 = _mm_loadu_si128((__m128i*)&p[0]); in SimpleVFilter16SSE2()
604 __m128i q1 = _mm_loadu_si128((__m128i*)&p[stride]); in SimpleVFilter16SSE2()
655 e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]); \
656 e2 = _mm_loadu_si128((__m128i*)&(p)[1 * stride]); \
657 e3 = _mm_loadu_si128((__m128i*)&(p)[2 * stride]); \
658 e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]); \
/external/neven/Embedded/common/src/b_BasicEm/
DMathSSE2.c269 m_XMM0 = _mm_loadu_si128( (__m128i *)&0[vec1L] ); in bbs_dotProduct_u128SSE2()
272 m_XMM2 = _mm_loadu_si128( (__m128i *)&0[vec2L] ); in bbs_dotProduct_u128SSE2()
274 m_XMM6 = _mm_loadu_si128( (__m128i *)&8[vec1L] ); in bbs_dotProduct_u128SSE2()
280 m_XMM3 = _mm_loadu_si128( (__m128i *)&8[vec2L] ); in bbs_dotProduct_u128SSE2()
292 m_XMM0 = _mm_loadu_si128( (__m128i *)&m_XMM5 ); in bbs_dotProduct_u128SSE2()
/external/libvpx/libvpx/vp9/encoder/x86/
Dvp9_dct32x32_sse2.c105 const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); in FDCT32x32_2D()
106 const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); in FDCT32x32_2D()
107 const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2)); in FDCT32x32_2D()
108 const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3)); in FDCT32x32_2D()
109 const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3)); in FDCT32x32_2D()
110 const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2)); in FDCT32x32_2D()
111 const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1)); in FDCT32x32_2D()
112 const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); in FDCT32x32_2D()
135 const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); in FDCT32x32_2D()
136 const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); in FDCT32x32_2D()
[all …]
/external/chromium_org/media/base/simd/
Dfilter_yuv_sse2.cc44 __m128i src0 = _mm_loadu_si128(src0_128); in FilterYUVRows_SSE2()
45 __m128i src1 = _mm_loadu_si128(src1_128); in FilterYUVRows_SSE2()
Dconvert_rgb_to_yuv_sse2.cc175 __m128i rgb_row_1 = _mm_loadu_si128( in ConvertRGB32ToYUVRow_SSE2()
215 __m128i rgb_row_2 = _mm_loadu_si128( in ConvertRGB32ToYUVRow_SSE2()
/external/libvpx/libvpx/vp9/common/x86/
Dvp9_loopfilter_intrin_avx2.c412 p4 = _mm_loadu_si128((__m128i *) (s - 5 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
413 p3 = _mm_loadu_si128((__m128i *) (s - 4 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
414 p2 = _mm_loadu_si128((__m128i *) (s - 3 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
415 p1 = _mm_loadu_si128((__m128i *) (s - 2 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
416 p0 = _mm_loadu_si128((__m128i *) (s - 1 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
417 q0 = _mm_loadu_si128((__m128i *) (s - 0 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
418 q1 = _mm_loadu_si128((__m128i *) (s + 1 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
419 q2 = _mm_loadu_si128((__m128i *) (s + 2 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
420 q3 = _mm_loadu_si128((__m128i *) (s + 3 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
421 q4 = _mm_loadu_si128((__m128i *) (s + 4 * p)); in mb_lpf_horizontal_edge_w_avx2_16()
[all …]
Dvp9_loopfilter_intrin_sse2.c405 p4 = _mm_loadu_si128((__m128i *)(s - 5 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
406 p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
407 p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
408 p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
409 p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
410 q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
411 q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
412 q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
413 q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
414 q4 = _mm_loadu_si128((__m128i *)(s + 4 * p)); in mb_lpf_horizontal_edge_w_sse2_16()
[all …]
/external/chromium_org/skia/ext/
Dconvolver_SSE2.cc60 __m128i src8 = _mm_loadu_si128(row_to_filter); in ConvolveHorizontally_SSE2()
111 __m128i src8 = _mm_loadu_si128(row_to_filter); in ConvolveHorizontally_SSE2()
190 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ in Convolve4RowsHorizontally_SSE2()
299 __m128i src8 = _mm_loadu_si128(src); in ConvolveVertically_SSE2()
382 __m128i src8 = _mm_loadu_si128(src); in ConvolveVertically_SSE2()
/external/libvpx/libvpx/vp8/encoder/x86/
Ddenoising_sse2.c53 const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0])); in vp8_denoiser_filter_sse2()
54 const __m128i v_mc_running_avg_y = _mm_loadu_si128( in vp8_denoiser_filter_sse2()
/external/skia/src/opts/
DSkBlitRow_opts_SSE2.cpp49 __m128i src_pixel = _mm_loadu_si128(s); in S32_Blend_BlitRow32_SSE2()
137 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
185 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
262 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Blend_BlitRow32_SSE2()
377 __m128i src_pixel = _mm_loadu_si128(s); in Color32_SSE2()
DSkBitmapFilter_opts_SSE2.cpp227 __m128i src8 = _mm_loadu_si128(row_to_filter); in convolveHorizontally_SSE2()
278 __m128i src8 = _mm_loadu_si128(row_to_filter); in convolveHorizontally_SSE2()
357 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ in convolve4RowsHorizontally_SSE2()
466 __m128i src8 = _mm_loadu_si128(src); in convolveVertically_SSE2()
549 __m128i src8 = _mm_loadu_si128(src); in convolveVertically_SSE2()
DSkBitmapProcState_opts_SSSE3.cpp50 const __m128i xx = _mm_loadu_si128(reinterpret_cast<const __m128i *>(xy)); in PrepareConstantsTwoPixelPairs()
96 _mm_loadu_si128(reinterpret_cast<const __m128i *>(xy)); in PrepareConstantsTwoPixelPairsDXDY()
/external/chromium_org/third_party/skia/src/opts/
DSkBlitRow_opts_SSE2.cpp49 __m128i src_pixel = _mm_loadu_si128(s); in S32_Blend_BlitRow32_SSE2()
137 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
185 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
262 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Blend_BlitRow32_SSE2()
377 __m128i src_pixel = _mm_loadu_si128(s); in Color32_SSE2()
DSkBitmapFilter_opts_SSE2.cpp227 __m128i src8 = _mm_loadu_si128(row_to_filter); in convolveHorizontally_SSE2()
278 __m128i src8 = _mm_loadu_si128(row_to_filter); in convolveHorizontally_SSE2()
357 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ in convolve4RowsHorizontally_SSE2()
466 __m128i src8 = _mm_loadu_si128(src); in convolveVertically_SSE2()
549 __m128i src8 = _mm_loadu_si128(src); in convolveVertically_SSE2()
DSkBitmapProcState_opts_SSSE3.cpp50 const __m128i xx = _mm_loadu_si128(reinterpret_cast<const __m128i *>(xy)); in PrepareConstantsTwoPixelPairs()
96 _mm_loadu_si128(reinterpret_cast<const __m128i *>(xy)); in PrepareConstantsTwoPixelPairsDXDY()
/external/chromium_org/third_party/angle/src/libGLESv2/renderer/
DImageSSE2.cpp40 __m128i sourceData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&source[x])); in loadRGBAUByteDataToBGRASSE2()
/external/pixman/test/
Dutils-prng.c88 _mm_storeu_si128 (addr, _mm_loadu_si128 ((__m128i *)d)); in store_rand_128_data()
/external/mesa3d/src/gallium/drivers/llvmpipe/
Dlp_setup_tri.c355 vertx = _mm_loadu_si128((__m128i *)position->x); /* vertex x coords */ in do_triangle_ccw()
356 verty = _mm_loadu_si128((__m128i *)position->y); /* vertex y coords */ in do_triangle_ccw()
/external/chromium_org/third_party/mesa/src/src/gallium/drivers/llvmpipe/
Dlp_setup_tri.c355 vertx = _mm_loadu_si128((__m128i *)position->x); /* vertex x coords */ in do_triangle_ccw()
356 verty = _mm_loadu_si128((__m128i *)position->y); /* vertex y coords */ in do_triangle_ccw()
/external/chromium_org/third_party/openssl/openssl/crypto/poly1305/
Dpoly1305_vec.c156 _mm_storeu_si128((xmmi *)(dst + 0), _mm_loadu_si128((xmmi *)(dst + offset + 0))); in poly1305_block_copy()
157 _mm_storeu_si128((xmmi *)(dst + 16), _mm_loadu_si128((xmmi *)(dst + offset + 16))); in poly1305_block_copy()
163 _mm_loadu_si128((xmmi *)(dst + offset))); in poly1305_block_copy()

12