/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 3x8inc-minmax-sse-dup.c | 104 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup() local 111 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup() 114 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup()
|
D | 3x8inc-minmax-sse2-dup.c | 104 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup() local 111 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup() 114 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup()
|
D | 4x8inc-minmax-sse2-dup.c | 120 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() local 128 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() 132 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8inc-minmax-sse-dup.c | 120 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() local 128 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() 132 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup()
|
D | 5x8inc-minmax-sse-dup.c | 136 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() local 145 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() 150 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup()
|
D | 5x8inc-minmax-sse2-dup.c | 136 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() local 145 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() 150 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 3x8-minmax-sse2-dup.c | 102 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup() local 109 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup() 112 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup()
|
D | 3x8-minmax-sse-dup.c | 102 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup() local 109 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup() 112 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup()
|
D | 4x8-minmax-sse2-dup.c | 118 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() local 126 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() 130 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 118 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() local 126 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() 130 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup()
|
D | 5x8-minmax-sse-dup.c | 134 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() local 143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() 148 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 134 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() local 143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() 148 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 3x8-minmax-sse-dup.c | 121 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup() local 128 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup() 131 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup()
|
D | 3x8-minmax-sse2-dup.c | 121 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup() local 128 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup() 131 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 140 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() local 148 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() 152 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup()
|
D | 4x8-minmax-sse2-dup.c | 140 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() local 148 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() 152 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8-minmax-sse2-dup.c | 159 …const __m128 va2c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va2), _MM_SHUFFLE(2, 2… in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() local 168 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() 173 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup()
|
D | 5x8-minmax-sse-dup.c | 159 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() local 168 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() 173 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup()
|