43   __m128i sum, ta, tb;  in inner_product_gint32_full_1_sse41()  local
50     tb = _mm_load_si128 ((__m128i *) (b + i));  in inner_product_gint32_full_1_sse41()
54             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_full_1_sse41()
57             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_full_1_sse41()
60     tb = _mm_load_si128 ((__m128i *) (b + i + 4));  in inner_product_gint32_full_1_sse41()
64             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_full_1_sse41()
67             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_full_1_sse41()
82   __m128i sum[2], ta, tb;  in inner_product_gint32_linear_1_sse41()  local
93     tb = _mm_load_si128 ((__m128i *) (c[0] + i));  in inner_product_gint32_linear_1_sse41()
95             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_linear_1_sse41()
97             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_linear_1_sse41()
99     tb = _mm_load_si128 ((__m128i *) (c[1] + i));  in inner_product_gint32_linear_1_sse41()
101             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_linear_1_sse41()
103             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_linear_1_sse41()
125   __m128i sum[4], ta, tb;  in inner_product_gint32_cubic_1_sse41()  local
138     tb = _mm_load_si128 ((__m128i *) (c[0] + i));  in inner_product_gint32_cubic_1_sse41()
140             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
142             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
144     tb = _mm_load_si128 ((__m128i *) (c[1] + i));  in inner_product_gint32_cubic_1_sse41()
146             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
148             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
150     tb = _mm_load_si128 ((__m128i *) (c[2] + i));  in inner_product_gint32_cubic_1_sse41()
152             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
154             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
156     tb = _mm_load_si128 ((__m128i *) (c[3] + i));  in inner_product_gint32_cubic_1_sse41()
158             _mm_unpacklo_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()
160             _mm_unpackhi_epi32 (tb, tb)));  in inner_product_gint32_cubic_1_sse41()