/external/XNNPACK/src/f32-gemm/gen/ |
D | 4x8-minmax-sse2-dup.c | 119 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() local 127 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() 131 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 119 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() local 127 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() 131 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup()
|
D | 5x8-minmax-sse-dup.c | 135 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() local 144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() 149 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 135 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() local 144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() 149 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 4x8-minmax-sse-dup.c | 141 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() local 149 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() 153 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup()
|
D | 4x8-minmax-sse2-dup.c | 141 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() local 149 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() 153 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8-minmax-sse-dup.c | 160 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() local 169 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() 174 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 160 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() local 169 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() 174 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 4x8inc-minmax-sse-dup.c | 121 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() local 129 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() 133 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup()
|
D | 4x8inc-minmax-sse2-dup.c | 121 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() local 129 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() 133 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8inc-minmax-sse2-dup.c | 137 …const __m128 va3c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va3), _MM_SHUFFLE(2, 2… in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() local 146 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() 151 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup()
|
D | 5x8inc-minmax-sse-dup.c | 137 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() local 146 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() 151 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup()
|