Home
last modified time | relevance | path

Searched refs:wasm_i32x4_shr (Results 1 – 25 of 77) sorted by relevance

1234

/external/XNNPACK/src/qs8-vaddc/gen/
Dminmax-wasmsimd-x32.c51 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
52 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
53 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
54 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
55 …const v128_t vremGHIJ = wasm_i32x4_add(wasm_v128_and(vaccGHIJ, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
56 …const v128_t vremKLMN = wasm_i32x4_add(wasm_v128_and(vaccKLMN, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
57 …const v128_t vremOPQR = wasm_i32x4_add(wasm_v128_and(vaccOPQR, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
58 …const v128_t vremSTUV = wasm_i32x4_add(wasm_v128_and(vaccSTUV, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
60 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
61 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
[all …]
Dminmax-wasmsimd-x24.c48 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
49 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
50 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
51 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
52 …const v128_t vremGHIJ = wasm_i32x4_add(wasm_v128_and(vaccGHIJ, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
53 …const v128_t vremKLMN = wasm_i32x4_add(wasm_v128_and(vaccKLMN, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
55 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
56 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
57 …vacc89AB = wasm_i32x4_sub(wasm_i32x4_shr(vacc89AB, vshift), wasm_i32x4_gt(vrem89AB, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
58 …vaccCDEF = wasm_i32x4_sub(wasm_i32x4_shr(vaccCDEF, vshift), wasm_i32x4_gt(vremCDEF, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
[all …]
Dminmax-wasmsimd-x16.c45 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
46 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
47 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
48 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
50 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
51 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
52 …vacc89AB = wasm_i32x4_sub(wasm_i32x4_shr(vacc89AB, vshift), wasm_i32x4_gt(vrem89AB, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
53 …vaccCDEF = wasm_i32x4_sub(wasm_i32x4_shr(vaccCDEF, vshift), wasm_i32x4_gt(vremCDEF, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
75 … vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(vacc0123, 31)); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
76 … vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(vacc4567, 31)); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
[all …]
Dminmax-wasmsimd-x8.c42 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
43 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
45 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
46 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
66 … vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(vacc0123, 31)); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
67 … vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(vacc4567, 31)); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
69 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
70 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8()
/external/XNNPACK/src/qs8-vadd/gen/
Dminmax-wasmsimd-x32.c64 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
65 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
66 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
67 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
68 …const v128_t vremGHIJ = wasm_i32x4_add(wasm_v128_and(vaccGHIJ, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
69 …const v128_t vremKLMN = wasm_i32x4_add(wasm_v128_and(vaccKLMN, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
70 …const v128_t vremOPQR = wasm_i32x4_add(wasm_v128_and(vaccOPQR, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
71 …const v128_t vremSTUV = wasm_i32x4_add(wasm_v128_and(vaccSTUV, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
73 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
74 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
[all …]
Dminmax-wasmsimd-x24.c58 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
59 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
60 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
61 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
62 …const v128_t vremGHIJ = wasm_i32x4_add(wasm_v128_and(vaccGHIJ, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
63 …const v128_t vremKLMN = wasm_i32x4_add(wasm_v128_and(vaccKLMN, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
65 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
66 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
67 …vacc89AB = wasm_i32x4_sub(wasm_i32x4_shr(vacc89AB, vshift), wasm_i32x4_gt(vrem89AB, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
68 …vaccCDEF = wasm_i32x4_sub(wasm_i32x4_shr(vaccCDEF, vshift), wasm_i32x4_gt(vremCDEF, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
[all …]
Dminmax-wasmsimd-x16.c52 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
53 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
54 …const v128_t vrem89AB = wasm_i32x4_add(wasm_v128_and(vacc89AB, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
55 …const v128_t vremCDEF = wasm_i32x4_add(wasm_v128_and(vaccCDEF, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
57 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
58 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
59 …vacc89AB = wasm_i32x4_sub(wasm_i32x4_shr(vacc89AB, vshift), wasm_i32x4_gt(vrem89AB, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
60 …vaccCDEF = wasm_i32x4_sub(wasm_i32x4_shr(vaccCDEF, vshift), wasm_i32x4_gt(vremCDEF, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
87 … vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(vacc0123, 31)); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
88 … vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(vacc4567, 31)); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
[all …]
Dminmax-wasmsimd-x8.c46 …const v128_t vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
47 …const v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(va… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
49 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
50 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
74 … vrem0123 = wasm_i32x4_add(wasm_v128_and(vacc0123, vremainder_mask), wasm_i32x4_shr(vacc0123, 31)); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
75 … vrem4567 = wasm_i32x4_add(wasm_v128_and(vacc4567, vremainder_mask), wasm_i32x4_shr(vacc4567, 31)); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
77 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vacc0123, vshift), wasm_i32x4_gt(vrem0123, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
78 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vacc4567, vshift), wasm_i32x4_gt(vrem4567, vremainder_thr… in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup24x9-minmax-wasmsimd-mul16.c264 const v128_t vsign0123 = wasm_i32x4_shr(vacc0123, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
265 const v128_t vsign4567 = wasm_i32x4_shr(vacc4567, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
266 const v128_t vsign89AB = wasm_i32x4_shr(vacc89AB, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
267 const v128_t vsignCDEF = wasm_i32x4_shr(vaccCDEF, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
268 const v128_t vsignGHIJ = wasm_i32x4_shr(vaccGHIJ, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
269 const v128_t vsignKLMN = wasm_i32x4_shr(vaccKLMN, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
307 …m0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0123, vremainder_mask), wasm_i32x4_shr(vq31prod0123, … in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
308 …m4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31prod4567, … in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
309 …m89AB = wasm_i32x4_add(wasm_v128_and(vq31prod89AB, vremainder_mask), wasm_i32x4_shr(vq31prod89AB, … in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
310 …mCDEF = wasm_i32x4_add(wasm_v128_and(vq31prodCDEF, vremainder_mask), wasm_i32x4_shr(vq31prodCDEF, … in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
[all …]
Dup16x9-minmax-wasmsimd-mul16.c217 const v128_t vsign0123 = wasm_i32x4_shr(vacc0123, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
218 const v128_t vsign4567 = wasm_i32x4_shr(vacc4567, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
219 const v128_t vsign89AB = wasm_i32x4_shr(vacc89AB, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
220 const v128_t vsignCDEF = wasm_i32x4_shr(vaccCDEF, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
248 …m0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0123, vremainder_mask), wasm_i32x4_shr(vq31prod0123, … in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
249 …m4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31prod4567, … in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
250 …m89AB = wasm_i32x4_add(wasm_v128_and(vq31prod89AB, vremainder_mask), wasm_i32x4_shr(vq31prod89AB, … in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
251 …mCDEF = wasm_i32x4_add(wasm_v128_and(vq31prodCDEF, vremainder_mask), wasm_i32x4_shr(vq31prodCDEF, … in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
255 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0123, vshift), wasm_i32x4_gt(vrem0123, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
256 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
[all …]
Dup8x9-minmax-wasmsimd-mul16.c170 const v128_t vsign0123 = wasm_i32x4_shr(vacc0123, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
171 const v128_t vsign4567 = wasm_i32x4_shr(vacc4567, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
189 …m0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0123, vremainder_mask), wasm_i32x4_shr(vq31prod0123, … in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
190 …m4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31prod4567, … in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
194 …vacc0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0123, vshift), wasm_i32x4_gt(vrem0123, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
195 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
286 const v128_t vsign0123 = wasm_i32x4_shr(vacc0123, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
287 const v128_t vsign4567 = wasm_i32x4_shr(vacc4567, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
305 …m0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0123, vremainder_mask), wasm_i32x4_shr(vq31prod0123, … in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
306 …m4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31prod4567, … in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
[all …]
/external/XNNPACK/src/f32-sigmoid/gen/
Dwasmsimd-p5-div-x24.c152 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
153 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
154 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
155 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
156 vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
157 vfKLMN = wasm_v128_bitselect(vfKLMN, wasm_f32x4_sub(vone, vfKLMN), wasm_i32x4_shr(vxKLMN, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
191 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
219 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
Dwasmsimd-p5-div-x20.c136 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
137 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
138 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
139 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
140 vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
173 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
201 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
Dwasmsimd-p5-div-x16.c120 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
121 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
122 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
123 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
155 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
183 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
Dwasmsimd-p5-div-x12.c104 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
105 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
106 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
137 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
165 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
/external/XNNPACK/src/qs8-vaddc/
Dwasmsimd.c.in45 … = wasm_i32x4_add(wasm_v128_and(vacc${ABC[N:N+4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[N:N+…
48 …vacc${ABC[N:N+4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[N:N+4]}, vshift), wasm_i32x4_gt(vrem${…
91 …]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[0:4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[0:4]…
92 …]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[4:8]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[4:8]…
94 …vacc${ABC[0:4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[0:4]}, vshift), wasm_i32x4_gt(vrem${ABC[…
95 …vacc${ABC[4:8]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[4:8]}, vshift), wasm_i32x4_gt(vrem${ABC[…
/external/XNNPACK/src/qs8-dwconv/
Dunipass-wasmsimd-mul16.c.in68 const v128_t vsign${ABC[C:C+4]} = wasm_i32x4_shr(vacc${ABC[C:C+4]}, 31);
84 …asm_i32x4_add(wasm_v128_and(vq31prod${ABC[C:C+4]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[…
89 …vacc${ABC[C:C+4]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[C:C+4]}, vshift), wasm_i32x4_gt(vr…
143 const v128_t vsign${ABC[0:4]} = wasm_i32x4_shr(vacc${ABC[0:4]}, 31);
144 const v128_t vsign${ABC[4:8]} = wasm_i32x4_shr(vacc${ABC[4:8]}, 31);
162 … wasm_i32x4_add(wasm_v128_and(vq31prod${ABC[0:4]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[…
163 … wasm_i32x4_add(wasm_v128_and(vq31prod${ABC[4:8]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[…
167 …vacc${ABC[0:4]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[0:4]}, vshift), wasm_i32x4_gt(vrem${…
168 …vacc${ABC[4:8]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[4:8]}, vshift), wasm_i32x4_gt(vrem${…
/external/XNNPACK/src/f32-velu/gen/
Dvelu-wasmsimd-arm-rr2-p6-x24.c158 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
160 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
162 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
164 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
166 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
168 const v128_t vsignmKLMN = wasm_i32x4_shr(vxKLMN, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
210 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
240 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
Dvelu-wasmsimd-arm-rr2-p6-x20.c142 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
144 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
146 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
148 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
150 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
190 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
220 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
Dvelu-wasmsimd-arm-rr2-p6-x16.c126 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
128 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
130 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
132 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
170 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
200 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
Dvelu-wasmsimd-x86-rr2-p6-x24.c176 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
178 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
180 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
182 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
184 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
186 const v128_t vsignmKLMN = wasm_i32x4_shr(vxKLMN, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
231 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
264 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-p6-x20.c157 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
159 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
161 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
163 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
165 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
208 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
241 const v128_t vsignm = wasm_i32x4_shr(vx, 31); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
/external/XNNPACK/src/qs8-requantization/
Dq31-wasmsimd.c104 wasm_i32x4_sub(wasm_i32x4_shr(x_q31product, shift), wasm_i32x4_gt(x_remainder, vthreshold)); in xnn_qs8_requantize_q31__wasmsimd()
106 wasm_i32x4_sub(wasm_i32x4_shr(y_q31product, shift), wasm_i32x4_gt(y_remainder, vthreshold)); in xnn_qs8_requantize_q31__wasmsimd()
108 wasm_i32x4_sub(wasm_i32x4_shr(z_q31product, shift), wasm_i32x4_gt(z_remainder, vthreshold)); in xnn_qs8_requantize_q31__wasmsimd()
110 wasm_i32x4_sub(wasm_i32x4_shr(w_q31product, shift), wasm_i32x4_gt(w_remainder, vthreshold)); in xnn_qs8_requantize_q31__wasmsimd()
/external/XNNPACK/src/qu8-requantization/
Dq31-wasmsimd.c104 wasm_i32x4_sub(wasm_i32x4_shr(x_q31product, shift), wasm_i32x4_gt(x_remainder, vthreshold)); in xnn_qu8_requantize_q31__wasmsimd()
106 wasm_i32x4_sub(wasm_i32x4_shr(y_q31product, shift), wasm_i32x4_gt(y_remainder, vthreshold)); in xnn_qu8_requantize_q31__wasmsimd()
108 wasm_i32x4_sub(wasm_i32x4_shr(z_q31product, shift), wasm_i32x4_gt(z_remainder, vthreshold)); in xnn_qu8_requantize_q31__wasmsimd()
110 wasm_i32x4_sub(wasm_i32x4_shr(w_q31product, shift), wasm_i32x4_gt(w_remainder, vthreshold)); in xnn_qu8_requantize_q31__wasmsimd()
/external/XNNPACK/src/qs8-vadd/
Dwasmsimd.c.in51 … = wasm_i32x4_add(wasm_v128_and(vacc${ABC[N:N+4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[N:N+…
54 …vacc${ABC[N:N+4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[N:N+4]}, vshift), wasm_i32x4_gt(vrem${…
102 …]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[0:4]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[0:4]…
103 …]} = wasm_i32x4_add(wasm_v128_and(vacc${ABC[4:8]}, vremainder_mask), wasm_i32x4_shr(vacc${ABC[4:8]…
105 …vacc${ABC[0:4]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[0:4]}, vshift), wasm_i32x4_gt(vrem${ABC[…
106 …vacc${ABC[4:8]} = wasm_i32x4_sub(wasm_i32x4_shr(vacc${ABC[4:8]}, vshift), wasm_i32x4_gt(vrem${ABC[…

1234