Lines Matching refs:vi
62 __m128 vi${M}x3012 = _mm_setzero_ps();
65 __m128 vi${M}x4567 = _mm_loadu_ps(i${M});
73 __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${M+K}x4567, vk${K}2));
75 __m128 vo${M}p${K} = _mm_mul_ps(vi${M+K}x4567, vk${K}2);
77 …vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x4567, vk${…
80 … const __m128 vi${M}x7456 = _mm_shuffle_ps(vi${M}x4567, vi${M}x4567, _MM_SHUFFLE(2, 1, 0, 3));
83 const __m128 vi${M}x89AB = _mm_loadu_ps(i${M});
87 const __m128 vi${M}x3456 = _mm_move_ss(vi${M}x7456, vi${M}x3012);
91 …vo${M}p${(K+5) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+5) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x34…
94 … const __m128 vi${M}x2345 = _mm_shuffle_ps(vi${M}x3012, vi${M}x7456, _MM_SHUFFLE(2, 1, 0, 3));
95 vi${M}x3012 = vi${M}x7456;
98 const __m128 vi${M}x8567 = _mm_move_ss(vi${M}x4567, vi${M}x89AB);
99 vi${M}x4567 = vi${M}x89AB;
103 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+10) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x2345, vk${K}0)…
106 … const __m128 vi${M}x5678 = _mm_shuffle_ps(vi${M}x8567, vi${M}x8567, _MM_SHUFFLE(0, 3, 2, 1));
110 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+15) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x5678, vk${K}3)…
113 … const __m128 vi${M}x6789 = _mm_shuffle_ps(vi${M}x5678, vi${M}x89AB, _MM_SHUFFLE(1, 0, 2, 1));
117 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+20) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x6789, vk${K}4)…
143 __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${M+K}x4567, vk${K}2));
145 __m128 vo${M}p${K} = _mm_mul_ps(vi${M+K}x4567, vk${K}2);
147 …vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x4567, vk${…
150 … const __m128 vi${M}x7456 = _mm_shuffle_ps(vi${M}x4567, vi${M}x4567, _MM_SHUFFLE(2, 1, 0, 3));
153 const __m128 vi${M}x89AB = _mm_and_ps(_mm_loadu_ps(i${M}), vmask);
157 const __m128 vi${M}x3456 = _mm_move_ss(vi${M}x7456, vi${M}x3012);
161 …vo${M}p${(K+5) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+5) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x34…
164 … const __m128 vi${M}x2345 = _mm_shuffle_ps(vi${M}x3012, vi${M}x7456, _MM_SHUFFLE(2, 1, 0, 3));
165 vi${M}x3012 = vi${M}x7456;
168 const __m128 vi${M}x8567 = _mm_move_ss(vi${M}x4567, vi${M}x89AB);
169 vi${M}x4567 = vi${M}x89AB;
173 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+10) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x2345, vk${K}0)…
176 … const __m128 vi${M}x5678 = _mm_shuffle_ps(vi${M}x8567, vi${M}x8567, _MM_SHUFFLE(0, 3, 2, 1));
180 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+15) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x5678, vk${K}3)…
183 … const __m128 vi${M}x6789 = _mm_shuffle_ps(vi${M}x5678, vi${M}x89AB, _MM_SHUFFLE(1, 0, 2, 1));
187 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+20) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x6789, vk${K}4)…
214 vi${M}x4567 = _mm_and_ps(vi${M}x4567, vmask);
219 __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${M+K}x4567, vk${K}2));
221 __m128 vo${M}p${K} = _mm_mul_ps(vi${M+K}x4567, vk${K}2);
223 …vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x4567, vk${…
226 … const __m128 vi${M}x7456 = _mm_shuffle_ps(vi${M}x4567, vi${M}x4567, _MM_SHUFFLE(2, 1, 0, 3));
229 const __m128 vi${M}x3456 = _mm_move_ss(vi${M}x7456, vi${M}x3012);
233 …vo${M}p${(K+5) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+5) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x34…
236 … const __m128 vi${M}x2345 = _mm_shuffle_ps(vi${M}x3012, vi${M}x7456, _MM_SHUFFLE(2, 1, 0, 3));
240 const __m128 vi${M}x8567 = _mm_move_ss(vi${M}x4567, vzero);
244 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+10) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x2345, vk${K}0)…
247 … const __m128 vi${M}x5678 = _mm_shuffle_ps(vi${M}x8567, vi${M}x8567, _MM_SHUFFLE(0, 3, 2, 1));
251 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+15) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x5678, vk${K}3)…
254 const __m128 vi${M}x6789 = _mm_shuffle_ps(vi${M}x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
258 …) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+20) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x6789, vk${K}4)…