Lines Matching refs:a128
3527 __m128i a128, b128, res; in vmul_s8() local
3528 a128 = _MM_CVTEPI8_EPI16 (_pM128i(a)); // SSE 4.1 use low 64 bits in vmul_s8()
3530 res = _mm_mullo_epi16 (a128, b128); in vmul_s8()
3556 __m128i mask, a128, b128, res; in vmul_u8() local
3558 a128 = _MM_CVTEPU8_EPI16 (_pM128i(a)); in vmul_u8()
3560 res = _mm_mullo_epi16 (a128, b128); in vmul_u8()
3703 __m128i low, hi, a128,b128; in vmull_s16()
3704 a128 = _pM128i(a); in vmull_s16()
3706 low = _mm_mullo_epi16(a128,b128); in vmull_s16()
3707 hi = _mm_mulhi_epi16(a128,b128); in vmull_s16()
3715 __m128i ab, ba, a128, b128; in vmull_s32() local
3716 a128 = _pM128i(a); in vmull_s32()
3718 ab = _mm_unpacklo_epi32 (a128, b128); //a0, b0, a1,b1 in vmull_s32()
3719 ba = _mm_unpacklo_epi32 (b128, a128); //b0, a0, b1,a1 in vmull_s32()
3742 __m128i a128,b128,low, hi; in vmull_u16()
3743 a128 = _pM128i(a); in vmull_u16()
3745 low = _mm_mullo_epi16(a128,b128); in vmull_u16()
3746 hi = _mm_mulhi_epu16(a128,b128); in vmull_u16()
3755 __m128i ab, ba, a128, b128; in vmull_u32() local
3756 a128 = _pM128i(a); in vmull_u32()
3758 ab = _mm_unpacklo_epi32 (a128, b128); //a0, b0, a1,b1 in vmull_u32()
3759 ba = _mm_unpacklo_epi32 (b128, a128); //b0, a0, b1,a1 in vmull_u32()
3767 __m128i a128,b128, c1, a128_16, bmasked_16, res, tmp, bmasked; in vmull_p8() local
3769 a128 = _pM128i(a); in vmull_p8()
3771 c1 = _mm_cmpeq_epi8 (a128,a128); //all ones 0xff.... in vmull_p8()
3775 a128_16 = _MM_CVTEPU8_EPI16 (a128); // SSE 4.1 in vmull_p8()
8214 __m128i a128, r128; variable
8215 a128 = _MM_CVTEPI8_EPI16 (_pM128i(a)); //SSE 4.1
8216 r128 = _mm_slli_epi16 (a128, b);
8226 __m128i a128, r128; variable
8227 a128 = _MM_CVTEPI16_EPI32 (_pM128i(a)); //SSE 4.1
8228 r128 = _mm_slli_epi32 (a128, b); //shift_res
8264 __m128i a128, r128; variable
8265 a128 = _MM_CVTEPU8_EPI16 (_pM128i(a)); //SSE 4.1
8266 r128 = _mm_slli_epi16 (a128, b); //shift_res
8276 __m128i a128, r128; variable
8277 a128 = _MM_CVTEPU16_EPI32 (_pM128i(a)); //SSE 4.1
8278 r128 = _mm_slli_epi32 (a128, b); //shift_res
8306 __m128i a128, r128_1, r128_2; variable
8307 a128 = _MM_CVTEPI8_EPI16 (a); //SSE 4.1
8308 r128_1 = _mm_slli_epi16 (a128, b);
8310 a128 = _mm_shuffle_epi32 (a, _SWAP_HI_LOW32);
8311 a128 = _MM_CVTEPI8_EPI16 (a128);
8312 r128_2 = _mm_slli_epi16 (a128, b);
8321 __m128i a128, r128_1, r128_2; variable
8322 a128 = _MM_CVTEPI16_EPI32 (a); //SSE 4.1
8323 r128_1 = _mm_slli_epi32 (a128, b); //shift_res
8325 a128 = _mm_shuffle_epi32 (a, _SWAP_HI_LOW32);
8326 a128 = _MM_CVTEPI16_EPI32 (a128);
8327 r128_2 = _mm_slli_epi32 (a128, b);
8375 __m128i a128, r128_1, r128_2; variable
8376 a128 = _MM_CVTEPU8_EPI16 (a); //SSE 4.1
8377 r128_1 = _mm_slli_epi16 (a128, b);
8379 a128 = _mm_shuffle_epi32 (a, _SWAP_HI_LOW32);
8380 a128 = _MM_CVTEPU8_EPI16 (a128);
8381 r128_2 = _mm_slli_epi16 (a128, b);
8435 __m128i a128, r128; variable
8436 a128 = _MM_CVTEPI8_EPI16 (_pM128i(a)); //SSE 4.1
8437 r128 = _mm_slli_epi16 (a128, b);
8446 __m128i a128, r128; variable
8447 a128 = _MM_CVTEPI16_EPI32 (_pM128i(a)); //SSE 4.1
8448 r128 = _mm_slli_epi32 (a128, b); //shift_res
8477 __m128i a128, r128_1, r128_2; variable
8478 a128 = _MM_CVTEPI8_EPI16 (a); //SSE 4.1
8479 r128_1 = _mm_slli_epi16 (a128, b);
8481 a128 = _mm_shuffle_epi32 (a, _SWAP_HI_LOW32);
8482 a128 = _MM_CVTEPI8_EPI16 (a128);
8483 r128_2 = _mm_slli_epi16 (a128, b);
8491 __m128i a128, r128_1, r128_2; variable
8492 a128 = _MM_CVTEPI16_EPI32 (a); //SSE 4.1
8493 r128_1 = _mm_slli_epi32 (a128, b); //shift_res
8495 a128 = _mm_shuffle_epi32 (a, _SWAP_HI_LOW32);
8496 a128 = _MM_CVTEPI16_EPI32 (a128);
8497 r128_2 = _mm_slli_epi32 (a128, b);