/external/XNNPACK/src/f32-gemm/gen/ |
D | 3x8-minmax-sse-dup.c | 116 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup() local 123 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup() 126 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse_dup()
|
D | 3x8-minmax-sse2-dup.c | 116 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup() local 123 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup() 126 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup()
|
D | 4x8-minmax-sse2-dup.c | 135 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() local 143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup() 147 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 135 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() local 143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup() 147 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_4x8__sse_dup()
|
D | 5x8-minmax-sse-dup.c | 154 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() local 163 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup() 168 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 154 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() local 163 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup() 168 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 3x8inc-minmax-sse-dup.c | 118 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup() local 125 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup() 128 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup()
|
D | 3x8inc-minmax-sse2-dup.c | 118 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup() local 125 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup() 128 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup()
|
D | 4x8inc-minmax-sse-dup.c | 137 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() local 145 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup() 149 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup()
|
D | 4x8inc-minmax-sse2-dup.c | 137 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() local 145 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup() 149 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8inc-minmax-sse2-dup.c | 156 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() local 165 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup() 170 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup()
|
D | 5x8inc-minmax-sse-dup.c | 156 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() local 165 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup() 170 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 3x8-minmax-sse-dup.c | 135 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup() local 142 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup() 145 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse_dup()
|
D | 3x8-minmax-sse2-dup.c | 135 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup() local 142 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup() 145 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_3x8__sse2_dup()
|
D | 4x8-minmax-sse-dup.c | 157 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() local 165 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup() 169 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse_dup()
|
D | 4x8-minmax-sse2-dup.c | 157 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() local 165 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup() 169 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_4x8__sse2_dup()
|
D | 5x8-minmax-sse-dup.c | 179 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() local 188 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup() 193 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse_dup()
|
D | 5x8-minmax-sse2-dup.c | 179 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() local 188 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup() 193 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__sse2_dup()
|