/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 4x8inc-minmax-sse2-dup.c | 138 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() local 146 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() 150 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8inc-minmax-sse-dup.c | 138 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() local 146 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() 150 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup()
|
D | 5x8inc-minmax-sse-dup.c | 157 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() local 166 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() 171 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup()
|
D | 5x8inc-minmax-sse2-dup.c | 157 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() local 166 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() 171 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 4x8-minmax-sse2-dup.c | 136 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() local 144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() 148 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 136 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() local 144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() 148 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup()
|
D | 5x8-minmax-sse-dup.c | 155 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() local 164 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() 169 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 155 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() local 164 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() 169 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 4x8-minmax-sse-dup.c | 158 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() local 166 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() 170 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup()
|
D | 4x8-minmax-sse2-dup.c | 158 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() local 166 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() 170 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8-minmax-sse2-dup.c | 180 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() local 189 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() 194 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup()
|
D | 5x8-minmax-sse-dup.c | 180 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() local 189 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() 194 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup()
|