Lines Matching refs:vi

75       __m128 vi${M}x7531 = _mm_setzero_ps();
80         const __m128 vi${M}x89AB = _mm_loadu_ps(i${M});
81         const __m128 vi${M}xCDEF = _mm_loadu_ps(i${M} + 4);
85 …      const __m128 vi${M}x8ACE = _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
86 …      const __m128 vi${M}x9BDF = _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
91             __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}1));
93             __m128 vo${M}p${K} = _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}1);
95 …vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x8ACE, vk…
98 …      const __m128 vi${M}xF9BD = _mm_shuffle_ps(vi${M}x9BDF, vi${M}x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
103             __m128 vo${M}p${K+3} = _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}2);
105 …3) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+3) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}…
108         const __m128 vi${M}x7BDF = _mm_move_ss(vi${M}xF9BD, vi${M}x7531);
111         vi${M}x7531 = vi${M}xF9BD;
115 …vo${M}p${(K+6) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+6) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x…
140         const __m128 vi${M}x89AB = _mm_loadu_ps(i${M});
141         const __m128 vi${M}xCDEF = _mm_loadu_ps(i${M} + 4);
144 …const __m128 vi${M}x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHU…
145 …const __m128 vi${M}x9BDF = _mm_and_ps(vmask_odd,  _mm_shuffle_ps(vi${M}x89AB, vi${M}xCDEF, _MM_SHU…
150             __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}1));
152             __m128 vo${M}p${K} = _mm_mul_ps(vi${2*M+K}x8ACE, vk${K}1);
154 …vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x8ACE, vk…
157 …      const __m128 vi${M}xF9BD = _mm_shuffle_ps(vi${M}x9BDF, vi${M}x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
162             __m128 vo${M}p${K+3} = _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}2);
164 …3) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+3) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x9BDF, vk${K}…
167         const __m128 vi${M}x7BDF = _mm_move_ss(vi${M}xF9BD, vi${M}x7531);
170         vi${M}x7531 = vi${M}xF9BD;
174 …vo${M}p${(K+6) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+6) % ACCUMULATORS}, _mm_mul_ps(vi${2*M+K}x…