/external/XNNPACK/src/f32-vrnd/gen/ |
D | vrndu-wasmsimd-addsub-x8.c | 39 const v128_t vrndmask0123 = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx0123)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 40 const v128_t vrndmask4567 = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx4567)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 48 const v128_t vadjmask0123 = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx0123, vrndx0123)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 49 const v128_t vadjmask4567 = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx4567, vrndx4567)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 66 const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 69 const v128_t vadjmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx, vrndx)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 80 const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8() 83 const v128_t vadjmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx, vrndx)); in xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8()
|
/external/XNNPACK/src/f32-spmm/gen/ |
D | 32x1-minmax-wasmsimd-x86.c | 73 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 74 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 75 v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 76 v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 77 v128_t voutGHIJ = wasm_v128_bitselect(vaccGHIJ, vmax, wasm_f32x4_le(vaccGHIJ, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 78 v128_t voutKLMN = wasm_v128_bitselect(vaccKLMN, vmax, wasm_f32x4_le(vaccKLMN, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 79 v128_t voutOPQR = wasm_v128_bitselect(vaccOPQR, vmax, wasm_f32x4_le(vaccOPQR, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 80 v128_t voutSTUV = wasm_v128_bitselect(vaccSTUV, vmax, wasm_f32x4_le(vaccSTUV, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 131 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() 132 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() [all …]
|
D | 32x1-minmax-wasmsimd-x86-pipelined.c | 87 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 88 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 89 v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 90 v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 91 v128_t voutGHIJ = wasm_v128_bitselect(vaccGHIJ, vmax, wasm_f32x4_le(vaccGHIJ, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 92 v128_t voutKLMN = wasm_v128_bitselect(vaccKLMN, vmax, wasm_f32x4_le(vaccKLMN, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 93 v128_t voutOPQR = wasm_v128_bitselect(vaccOPQR, vmax, wasm_f32x4_le(vaccOPQR, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 94 v128_t voutSTUV = wasm_v128_bitselect(vaccSTUV, vmax, wasm_f32x4_le(vaccSTUV, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 145 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() 146 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() [all …]
|
D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 127 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 128 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 129 v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 130 v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 131 v128_t voutGHIJ = wasm_v128_bitselect(vaccGHIJ, vmax, wasm_f32x4_le(vaccGHIJ, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 132 v128_t voutKLMN = wasm_v128_bitselect(vaccKLMN, vmax, wasm_f32x4_le(vaccKLMN, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 133 v128_t voutOPQR = wasm_v128_bitselect(vaccOPQR, vmax, wasm_f32x4_le(vaccOPQR, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 134 v128_t voutSTUV = wasm_v128_bitselect(vaccSTUV, vmax, wasm_f32x4_le(vaccSTUV, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 185 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() 186 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() [all …]
|
D | 32x1-minmax-wasmsimd-x86-x2.c | 142 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 143 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 144 v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 145 v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 146 v128_t voutGHIJ = wasm_v128_bitselect(vaccGHIJ, vmax, wasm_f32x4_le(vaccGHIJ, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 147 v128_t voutKLMN = wasm_v128_bitselect(vaccKLMN, vmax, wasm_f32x4_le(vaccKLMN, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 148 v128_t voutOPQR = wasm_v128_bitselect(vaccOPQR, vmax, wasm_f32x4_le(vaccOPQR, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 149 v128_t voutSTUV = wasm_v128_bitselect(vaccSTUV, vmax, wasm_f32x4_le(vaccSTUV, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 200 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() 201 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() [all …]
|
/external/XNNPACK/src/f32-vbinary/gen/ |
D | vmaxc-wasmsimd-x86-x16.c | 39 const v128_t vm0123 = wasm_f32x4_le(va0123, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16() 40 const v128_t vm4567 = wasm_f32x4_le(va4567, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16() 41 const v128_t vm89AB = wasm_f32x4_le(va89AB, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16() 42 const v128_t vmCDEF = wasm_f32x4_le(vaCDEF, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16() 60 const v128_t vm = wasm_f32x4_le(va, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16() 70 const v128_t vm = wasm_f32x4_le(va, vb); in xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16()
|
D | vmax-wasmsimd-x86-x16.c | 45 const v128_t vm0123 = wasm_f32x4_le(va0123, vb0123); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 46 const v128_t vm4567 = wasm_f32x4_le(va4567, vb4567); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 47 const v128_t vm89AB = wasm_f32x4_le(va89AB, vb89AB); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 48 const v128_t vmCDEF = wasm_f32x4_le(vaCDEF, vbCDEF); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 69 const v128_t vm = wasm_f32x4_le(va, vb); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 80 const v128_t vm = wasm_f32x4_le(va, vb); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16()
|
D | vmulc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vrsubc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vsubc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vrdivc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vdivc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vaddc-minmax-wasmsimd-x86-x16.c | 52 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16() 54 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16() 56 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16() 58 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16() 79 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16() 92 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16()
|
D | vsub-minmax-wasmsimd-x86-x16.c | 58 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16() 60 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16() 62 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16() 64 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16() 88 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16() 102 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16()
|
D | vmul-minmax-wasmsimd-x86-x16.c | 58 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16() 60 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16() 62 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16() 64 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16() 88 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16() 102 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16()
|
D | vadd-minmax-wasmsimd-x86-x16.c | 58 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16() 60 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16() 62 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16() 64 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16() 88 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16() 102 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16()
|
D | vdiv-minmax-wasmsimd-x86-x16.c | 58 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16() 60 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16() 62 const v128_t vngtmask89AB = wasm_f32x4_le(vy89AB, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16() 64 const v128_t vngtmaskCDEF = wasm_f32x4_le(vyCDEF, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16() 88 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16() 102 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max); in xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16()
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 6x8-minmax-wasmsimd-x86-loadsplat.c | 136 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 137 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 138 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 139 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 140 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 141 vacc5x0123 = wasm_v128_bitselect(vacc5x0123, vmax, wasm_f32x4_le(vacc5x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 142 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 143 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 144 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 145 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() [all …]
|
D | 5x8-minmax-wasmsimd-x86-loadsplat.c | 122 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 123 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 124 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 125 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 126 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 127 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 128 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 129 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 130 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 131 vacc4x4567 = wasm_v128_bitselect(vacc4x4567, vmax, wasm_f32x4_le(vacc4x4567, vmax)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 6x8inc-minmax-wasmsimd-x86-loadsplat.c | 138 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 139 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 140 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 141 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 142 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 143 vacc5x0123 = wasm_v128_bitselect(vacc5x0123, vmax, wasm_f32x4_le(vacc5x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 144 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 145 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 146 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 147 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() [all …]
|
D | 5x8inc-minmax-wasmsimd-x86-loadsplat.c | 124 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 125 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 126 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 127 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 128 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 129 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 130 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 131 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 132 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 133 vacc4x4567 = wasm_v128_bitselect(vacc4x4567, vmax, wasm_f32x4_le(vacc4x4567, vmax)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 6x8-minmax-wasmsimd-x86-loadsplat.c | 165 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 166 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 167 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 168 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 169 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 170 vacc5x0123 = wasm_v128_bitselect(vacc5x0123, vmax, wasm_f32x4_le(vacc5x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 171 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 172 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 173 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() 174 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() [all …]
|
D | 5x8-minmax-wasmsimd-x86-loadsplat.c | 148 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 149 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 150 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 151 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 152 vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 153 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 154 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 155 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 156 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() 157 vacc4x4567 = wasm_v128_bitselect(vacc4x4567, vmax, wasm_f32x4_le(vacc4x4567, vmax)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
|
/external/XNNPACK/src/f32-vmulcaddc/gen/ |
D | c8-minmax-wasmsimd-x86-2x.c | 73 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 74 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 75 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 76 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 103 vacc0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 104 vacc1 = wasm_v128_bitselect(vacc1, vmax, wasm_f32x4_le(vacc1, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 129 vacc0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() 130 vacc1 = wasm_v128_bitselect(vacc1, vmax, wasm_f32x4_le(vacc1, vmax)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
|
/external/XNNPACK/src/f32-ppmm/gen/ |
D | 4x8-minmax-wasmsimd-x86-splat.c | 96 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 97 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 98 vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 99 vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 100 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 101 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 102 vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 103 vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat()
|