| /external/XNNPACK/src/f16-f32-vcvt/gen/ |
| D | vcvt-scalar-x3.c | 61 …const uint32_t vdenorm2 = float_as_uint32(uint32_as_float((v2w2 >> 17) | vmagic_mask) - vmagic_bia… in xnn_f16_f32_vcvt_ukernel__scalar_x3() local
65 const uint32_t vf2 = vsign2 | (XNN_UNPREDICTABLE(v2w2 < vdenorm_cutoff) ? vdenorm2 : vnorm2); in xnn_f16_f32_vcvt_ukernel__scalar_x3()
|
| D | vcvt-scalar-x4.c | 66 …const uint32_t vdenorm2 = float_as_uint32(uint32_as_float((v2w2 >> 17) | vmagic_mask) - vmagic_bia… in xnn_f16_f32_vcvt_ukernel__scalar_x4() local
71 const uint32_t vf2 = vsign2 | (XNN_UNPREDICTABLE(v2w2 < vdenorm_cutoff) ? vdenorm2 : vnorm2); in xnn_f16_f32_vcvt_ukernel__scalar_x4()
|
| D | vcvt-wasmsimd-int16-x16.c | 60 …const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign1, vmagic_mask, 0, 8, 1, 9, 2,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local
77 wasm_v128_bitselect(vnorm2, vdenorm2, vxmask2)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16()
|
| D | vcvt-sse41-int16-x16.c | 60 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local
71 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16()
|
| D | vcvt-neon-int32-x16.c | 62 …const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign2, 1… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
72 …st uint32x4_t vf2 = vorrq_u32(vsign2, vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
|
| D | vcvt-avx-int16-x16.c | 60 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local
71 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x16()
|
| D | vcvt-neon-int16-x16.c | 56 …const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(v… in xnn_f16_f32_vcvt_ukernel__neon_int16_x16() local
67 vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x16()
|
| D | vcvt-wasmsimd-int32-x16.c | 64 …const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign2, 16), vmagic_bias), v… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16() local
74 const v128_t vf2 = wasm_v128_or(vsign2, wasm_v128_bitselect(vnorm2, vdenorm2, vxmask2)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
|
| D | vcvt-avx-int32-x16.c | 63 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
73 const __m128i vf2 = _mm_or_si128(vsign2, _mm_blendv_epi8(vdenorm2, vnorm2, vmask2)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
|
| D | vcvt-sse41-int32-x16.c | 63 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
73 const __m128i vf2 = _mm_or_si128(vsign2, _mm_blendv_epi8(vdenorm2, vnorm2, vmask2)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
|
| D | vcvt-wasmsimd-int16-x24.c | 67 …const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign1, vmagic_mask, 0, 8, 1, 9, 2,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24() local
89 wasm_v128_bitselect(vnorm2, vdenorm2, vxmask2)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24()
|
| D | vcvt-sse41-int16-x24.c | 67 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() local
81 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24()
|
| D | vcvt-sse2-int16-x16.c | 60 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16() local
74 _mm_or_si128(_mm_and_si128(vxmask2, vnorm2), _mm_andnot_si128(vxmask2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16()
|
| D | vcvt-sse2-int32-x16.c | 63 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16() local
76 _mm_or_si128(_mm_and_si128(vmask2, vnorm2), _mm_andnot_si128(vmask2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
|
| D | vcvt-avx-int16-x24.c | 67 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() local
81 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x24()
|
| D | vcvt-wasmsimd-int32-x24.c | 73 …const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign2, 16), vmagic_bias), v… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24() local
87 const v128_t vf2 = wasm_v128_or(vsign2, wasm_v128_bitselect(vnorm2, vdenorm2, vxmask2)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
|
| D | vcvt-sse2-int16-x24.c | 67 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24() local
84 _mm_or_si128(_mm_and_si128(vxmask2, vnorm2), _mm_andnot_si128(vxmask2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24()
|
| D | vcvt-neon-int16-x24.c | 62 …const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(v… in xnn_f16_f32_vcvt_ukernel__neon_int16_x24() local
76 vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x24()
|
| D | vcvt-sse41-int32-x24.c | 72 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24() local
86 const __m128i vf2 = _mm_or_si128(vsign2, _mm_blendv_epi8(vdenorm2, vnorm2, vmask2)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
|
| D | vcvt-neon-int32-x24.c | 71 …const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign2, 1… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24() local
85 …st uint32x4_t vf2 = vorrq_u32(vsign2, vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
|
| D | vcvt-wasmsimd-int16-x32.c | 74 …const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign1, vmagic_mask, 0, 8, 1, 9, 2,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32() local
101 wasm_v128_bitselect(vnorm2, vdenorm2, vxmask2)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32()
|
| D | vcvt-avx-int32-x24.c | 72 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__avx_int32_x24() local
86 const __m128i vf2 = _mm_or_si128(vsign2, _mm_blendv_epi8(vdenorm2, vnorm2, vmask2)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
|
| D | vcvt-sse41-int16-x32.c | 74 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32() local
91 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32()
|
| D | vcvt-avx-int16-x32.c | 74 …const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1… in xnn_f16_f32_vcvt_ukernel__avx_int16_x32() local
91 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x32()
|
| D | vcvt-neon-int16-x32.c | 68 …const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(v… in xnn_f16_f32_vcvt_ukernel__neon_int16_x32() local
85 vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x32()
|