| /external/XNNPACK/src/f16-f32-vcvt/gen/ |
| D | vcvt-wasmsimd-int16-x24.c | 69 …const v128_t vdenorm4 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign2, vmagic_mask, 0, 8, 1, 9, 2,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24() local
93 wasm_v128_bitselect(vnorm4, vdenorm4, vxmask4)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24()
|
| D | vcvt-sse41-int16-x24.c | 69 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() local
85 _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24()
|
| D | vcvt-avx-int16-x24.c | 69 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() local
85 _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x24()
|
| D | vcvt-wasmsimd-int32-x24.c | 75 …const v128_t vdenorm4 = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign4, 16), vmagic_bias), v… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24() local
89 const v128_t vf4 = wasm_v128_or(vsign4, wasm_v128_bitselect(vnorm4, vdenorm4, vxmask4)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
|
| D | vcvt-sse2-int16-x24.c | 69 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24() local
90 _mm_or_si128(_mm_and_si128(vxmask4, vnorm4), _mm_andnot_si128(vxmask4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24()
|
| D | vcvt-neon-int16-x24.c | 64 …const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(v… in xnn_f16_f32_vcvt_ukernel__neon_int16_x24() local
79 vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x24()
|
| D | vcvt-sse41-int32-x24.c | 74 …const __m128i vdenorm4 = _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
88 const __m128i vf4 = _mm_or_si128(vsign4, _mm_blendv_epi8(vdenorm4, vnorm4, vmask4)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
|
| D | vcvt-neon-int32-x24.c | 73 …const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign4, 1… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24() local
87 …st uint32x4_t vf4 = vorrq_u32(vsign4, vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
|
| D | vcvt-wasmsimd-int16-x32.c | 76 …const v128_t vdenorm4 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign2, vmagic_mask, 0, 8, 1, 9, 2,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32() local
105 wasm_v128_bitselect(vnorm4, vdenorm4, vxmask4)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32()
|
| D | vcvt-avx-int32-x24.c | 74 …const __m128i vdenorm4 = _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
88 const __m128i vf4 = _mm_or_si128(vsign4, _mm_blendv_epi8(vdenorm4, vnorm4, vmask4)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
|
| D | vcvt-sse41-int16-x32.c | 76 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32() local
95 _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32()
|
| D | vcvt-avx-int16-x32.c | 76 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__avx_int16_x32() local
95 _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x32()
|
| D | vcvt-neon-int16-x32.c | 70 …const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(v… in xnn_f16_f32_vcvt_ukernel__neon_int16_x32() local
88 vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x32()
|
| D | vcvt-sse2-int32-x24.c | 74 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24() local
93 _mm_or_si128(_mm_and_si128(vmask4, vnorm4), _mm_andnot_si128(vmask4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
|
| D | vcvt-neon-int32-x32.c | 82 …const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign4, 1… in xnn_f16_f32_vcvt_ukernel__neon_int32_x32() local
100 …st uint32x4_t vf4 = vorrq_u32(vsign4, vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
|
| D | vcvt-wasmsimd-int32-x32.c | 84 …const v128_t vdenorm4 = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign4, 16), vmagic_bias), v… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32() local
102 const v128_t vf4 = wasm_v128_or(vsign4, wasm_v128_bitselect(vnorm4, vdenorm4, vxmask4)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
|
| D | vcvt-avx-int32-x32.c | 83 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__avx_int32_x32() local
101 const __m128i vf4 = _mm_or_si128(vsign4, _mm_blendv_epi8(vdenorm4, vnorm4, vmask4)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
|
| D | vcvt-sse2-int16-x32.c | 76 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() local
100 _mm_or_si128(_mm_and_si128(vxmask4, vnorm4), _mm_andnot_si128(vxmask4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32()
|
| D | vcvt-sse41-int32-x32.c | 83 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x32() local
101 const __m128i vf4 = _mm_or_si128(vsign4, _mm_blendv_epi8(vdenorm4, vnorm4, vmask4)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x32()
|
| D | vcvt-sse2-int32-x32.c | 83 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x32() local
106 _mm_or_si128(_mm_and_si128(vmask4, vnorm4), _mm_andnot_si128(vmask4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x32()
|
| /external/XNNPACK/src/amalgam/ |
| D | sse2.c | 95 …const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() local
119 _mm_or_si128(_mm_and_si128(vxmask4, vnorm4), _mm_andnot_si128(vxmask4, vdenorm4))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32()
|