| /external/XNNPACK/src/f16-f32-vcvt/gen/ |
| D | vcvt-neon-int32-x8.c | 46 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local
49 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
52 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
57 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
73 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local
76 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
79 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
87 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x8()
|
| D | vcvt-wasmsimd-int32-x8.c | 48 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local
51 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
54 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
57 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
79 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local
82 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
85 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
94 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8()
|
| D | vcvt-avx-int32-x8.c | 47 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local
50 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
53 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
58 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
75 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local
78 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
81 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
90 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x8()
|
| D | vcvt-sse41-int32-x8.c | 47 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local
50 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
53 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
58 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
75 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local
78 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
81 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
90 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8()
|
| D | vcvt-sse2-int32-x8.c | 47 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local
50 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
53 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
59 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
77 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local
80 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
83 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
93 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8()
|
| D | vcvt-neon-int32-x16.c | 90 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
93 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
96 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
101 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
117 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
120 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
123 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
131 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x16()
|
| D | vcvt-wasmsimd-int32-x16.c | 95 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16() local
98 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
101 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
104 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
126 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16() local
129 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
132 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
141 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x16()
|
| D | vcvt-avx-int32-x16.c | 93 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
96 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
99 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
104 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
121 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
124 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
127 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
136 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x16()
|
| D | vcvt-sse41-int32-x16.c | 93 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
96 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
99 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
104 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
121 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
124 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
127 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
136 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16()
|
| D | vcvt-sse2-int32-x16.c | 97 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16() local
100 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
103 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
109 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
127 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16() local
130 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
133 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
143 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16()
|
| D | vcvt-wasmsimd-int32-x24.c | 112 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24() local
115 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
118 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
121 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
143 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24() local
146 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
149 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
158 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x24()
|
| D | vcvt-sse41-int32-x24.c | 110 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24() local
113 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
116 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
121 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
138 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24() local
141 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
144 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
153 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse41_int32_x24()
|
| D | vcvt-neon-int32-x24.c | 107 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24() local
110 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
113 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
118 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
134 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24() local
137 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
140 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
148 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x24()
|
| D | vcvt-avx-int32-x24.c | 110 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24() local
113 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
116 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
121 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
138 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24() local
141 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
144 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
153 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x24()
|
| D | vcvt-sse2-int32-x24.c | 116 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24() local
119 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
122 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
128 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
146 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24() local
149 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
152 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
162 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__sse2_int32_x24()
|
| D | vcvt-neon-int32-x32.c | 124 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x32() local
127 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
130 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
135 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
151 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__neon_int32_x32() local
154 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
157 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter… in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
165 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__neon_int32_x32()
|
| D | vcvt-wasmsimd-int32-x32.c | 129 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32() local
132 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
135 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
138 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
160 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32() local
163 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
166 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
175 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x32()
|
| D | vcvt-avx-int32-x32.c | 127 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x32() local
130 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
133 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
138 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
155 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_f16_f32_vcvt_ukernel__avx_int32_x32() local
158 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
161 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi… in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
170 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_f16_f32_vcvt_ukernel__avx_int32_x32()
|
| /external/XNNPACK/src/f16-f32-vcvt/ |
| D | neon-int32.c.in | 76 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); variable
79 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal…
82 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter…
87 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
103 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); variable
106 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal…
109 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinter…
117 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
|
| D | wasmsimd-int32.c.in | 83 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); variable
86 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)…
89 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias…
92 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff);
114 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); variable
117 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)…
120 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias…
129 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff);
|
| D | sse-int32.c.in | 89 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
92 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse…
95 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi…
104 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
125 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
128 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse…
131 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bi…
144 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
|
| /external/XNNPACK/src/math/ |
| D | cvt-f16-f32-wasmsimd-int32.c | 42 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi); in xnn_math_f16_f32_cvt__wasmsimd_int32() local
45 …const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset)… in xnn_math_f16_f32_cvt__wasmsimd_int32()
48 …const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_mask… in xnn_math_f16_f32_cvt__wasmsimd_int32()
51 const v128_t vmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff); in xnn_math_f16_f32_cvt__wasmsimd_int32()
|
| D | cvt-f16-f32-sse41-int32.c | 41 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_math_f16_f32_cvt__sse41_int32() local
44 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_math_f16_f32_cvt__sse41_int32()
47 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_ma… in xnn_math_f16_f32_cvt__sse41_int32()
50 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_math_f16_f32_cvt__sse41_int32()
|
| D | cvt-f16-f32-neon-int32.c | 40 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi); in xnn_math_f16_f32_cvt__neon_int32() local
43 …_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scal… in xnn_math_f16_f32_cvt__neon_int32()
46 … vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_mask, vnonsign_hi, 16)), vmagic_b… in xnn_math_f16_f32_cvt__neon_int32()
49 const uint32x4_t vmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff); in xnn_math_f16_f32_cvt__neon_int32()
|
| D | cvt-f16-f32-sse2-int32.c | 41 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi); in xnn_math_f16_f32_cvt__sse2_int32() local
44 …_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offse… in xnn_math_f16_f32_cvt__sse2_int32()
47 …m_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_ma… in xnn_math_f16_f32_cvt__sse2_int32()
50 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff); in xnn_math_f16_f32_cvt__sse2_int32()
|