| /external/XNNPACK/src/f16-f32-vcvt/gen/ |
| D | vcvt-neon-int32-x8.c | 51 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local
55 …4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
78 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local
82 …32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
|
| D | vcvt-wasmsimd-int16-x8.c | 51 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0, 8, 1, 9, 2… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8() local
59 wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8()
84 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0, 8, 1, 9, 2… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8() local
92 wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8()
|
| D | vcvt-neon-int16-x8.c | 47 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16… in xnn_f16_f32_vcvt_ukernel__neon_int16_x8() local
54 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x8()
74 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16… in xnn_f16_f32_vcvt_ukernel__neon_int16_x8() local
81 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x8()
|
| D | vcvt-sse2-int16-x8.c | 51 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8() local
58 _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8()
81 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8() local
88 _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8()
|
| D | vcvt-avx-int16-x8.c | 51 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__avx_int16_x8() local
57 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x8()
79 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__avx_int16_x8() local
85 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x8()
|
| D | vcvt-sse41-int16-x8.c | 51 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8() local
57 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8()
79 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8() local
85 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8()
|
| D | vcvt-wasmsimd-int32-x8.c | 53 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local
59 … const v128_t vf_lo = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
84 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local
88 v128_t vf = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
|
| D | vcvt-avx-int32-x8.c | 52 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local
56 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
80 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local
84 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
|
| D | vcvt-sse41-int32-x8.c | 52 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local
56 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
80 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local
84 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
|
| D | vcvt-sse2-int32-x8.c | 52 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local
57 _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
82 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local
87 _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
|
| D | vcvt-wasmsimd-int16-x16.c | 101 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0, 8, 1, 9, 2… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local
109 wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16()
134 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0, 8, 1, 9, 2… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local
142 wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16()
|
| D | vcvt-sse41-int16-x16.c | 95 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local
101 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16()
123 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local
129 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16()
|
| D | vcvt-neon-int32-x16.c | 95 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
99 …4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
122 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
126 …32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
|
| D | vcvt-avx-int16-x16.c | 95 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local
101 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x16()
123 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local
129 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_f16_f32_vcvt_ukernel__avx_int16_x16()
|
| D | vcvt-neon-int16-x16.c | 92 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16… in xnn_f16_f32_vcvt_ukernel__neon_int16_x16() local
99 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x16()
119 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16… in xnn_f16_f32_vcvt_ukernel__neon_int16_x16() local
126 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); in xnn_f16_f32_vcvt_ukernel__neon_int16_x16()
|
| D | vcvt-wasmsimd-int32-x16.c | 100 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16() local
106 … const v128_t vf_lo = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
131 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16() local
135 v128_t vf = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo)); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
|
| D | vcvt-avx-int32-x16.c | 98 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
102 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
126 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
130 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
|
| D | vcvt-sse41-int32-x16.c | 98 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
102 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
126 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
130 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo)); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
|
| /external/XNNPACK/src/f16-f32-vcvt/ |
| D | sse-int32.c.in | 94 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3…
99 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
102 _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
130 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi3…
135 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
138 _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));
|
| D | sse-int16.c.in | 102 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig…
109 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));
113 _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo)));
140 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig…
147 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));
151 _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo)));
|
| D | neon-int32.c.in | 81 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… variable
85 …4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
108 …const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_l… variable
112 …32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
|
| /external/XNNPACK/src/math/ |
| D | cvt-f16-f32-wasmsimd-int16.c | 44 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0, 8, 1, 9, 2… in xnn_math_f16_f32_cvt__wasmsimd_int16() local
53 wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vmask_lo)); in xnn_math_f16_f32_cvt__wasmsimd_int16()
|
| D | cvt-f16-f32-wasmsimd-int32.c | 47 …const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_mask… in xnn_math_f16_f32_cvt__wasmsimd_int32() local
53 … const v128_t vf_lo = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vmask_lo)); in xnn_math_f16_f32_cvt__wasmsimd_int32()
|
| D | cvt-f16-f32-sse41-int16.c | 44 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_math_f16_f32_cvt__sse41_int16() local
49 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask))); in xnn_math_f16_f32_cvt__sse41_int16()
|
| D | cvt-f16-f32-sse2-int16.c | 44 …const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsig… in xnn_math_f16_f32_cvt__sse2_int16() local
52 _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo))); in xnn_math_f16_f32_cvt__sse2_int16()
|