/external/XNNPACK/src/f32-vbinary/gen/ |
D | vmax-wasmsimd-x86-x16.c | 48 const v128_t vmCDEF = wasm_f32x4_le(vaCDEF, vbCDEF); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() local 53 v128_t vyCDEF = wasm_v128_bitselect(vbCDEF, vaCDEF, vmCDEF); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16()
|
D | vmin-wasmsimd-x86-x16.c | 48 const v128_t vmCDEF = wasm_f32x4_lt(vaCDEF, vbCDEF); in xnn_f32_vmin_ukernel__wasmsimd_x86_x16() local 53 v128_t vyCDEF = wasm_v128_bitselect(vaCDEF, vbCDEF, vmCDEF); in xnn_f32_vmin_ukernel__wasmsimd_x86_x16()
|
/external/XNNPACK/src/f32-vsigmoid/gen/ |
D | vsigmoid-sse2-rr2-p5-div-x16.c | 123 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x16() local 128 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x16()
|
D | vsigmoid-sse2-rr2-p5-div-x20.c | 139 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20() local 145 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20()
|
D | vsigmoid-sse2-rr2-p5-div-x24.c | 155 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24() local 162 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24()
|
D | vsigmoid-neonfma-rr1-p5-div-x16.c | 116 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16() local 121 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c | 131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 136 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x16.c | 131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() local 136 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x16.c | 131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 136 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-div-x20.c | 131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20() local 137 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
D | vsigmoid-sse2-rr2-lut64-p2-div-x16.c | 197 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x16() local 202 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x16()
|
D | vsigmoid-neonfma-rr1-p5-div-x24.c | 146 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24() local 153 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
D | vsigmoid-neonfma-rr1-lut64-p2-div-x16.c | 141 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local 146 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
D | vsigmoid-neon-rr2-p5-nr2recps-x16.c | 137 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local 142 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-div-x16.c | 135 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local 140 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x16.c | 150 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 155 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x20.c | 149 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 155 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse2-rr2-p6-x16.c | 131 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_velu_ukernel__sse2_rr2_p6_x16() local 137 const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
|
D | velu-sse2-rr2-p6-x20.c | 147 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_velu_ukernel__sse2_rr2_p6_x20() local 155 const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
|
D | velu-sse2-rr2-p6-x24.c | 163 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_velu_ukernel__sse2_rr2_p6_x24() local 173 const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
|
D | velu-neonfma-rr1-p6-x16.c | 123 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16() local 129 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
|
D | velu-neon-rr2-p6-x16.c | 129 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x16() local 135 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
|
D | velu-sse2-rr2-lut16-p3-x16.c | 208 …const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDE… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 214 const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF)); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16()
|
D | velu-neonfma-rr1-p6-x20.c | 138 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local 146 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
|
D | velu-neon-rr2-p6-x20.c | 145 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local 153 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
|