| /external/XNNPACK/src/f16-f32-vcvt/gen/ |
| D | vcvt-avx-int16-x8.c | 48 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x8() local 76 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x8() local
|
| D | vcvt-sse41-int16-x8.c | 48 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8() local 76 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x8() local
|
| D | vcvt-neon-int16-x8.c | 44 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x8() local 71 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x8() local
|
| D | vcvt-sse2-int16-x8.c | 48 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8() local 78 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x8() local
|
| D | vcvt-neon-int32-x8.c | 48 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local 75 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… in xnn_f16_f32_vcvt_ukernel__neon_int32_x8() local
|
| D | vcvt-sse41-int32-x8.c | 49 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local 77 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x8() local
|
| D | vcvt-wasmsimd-int16-x8.c | 48 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8() local 81 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x8() local
|
| D | vcvt-sse2-int32-x8.c | 49 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local 79 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x8() local
|
| D | vcvt-avx-int32-x8.c | 49 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local 77 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__avx_int32_x8() local
|
| D | vcvt-wasmsimd-int32-x8.c | 50 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local 81 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset)… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8() local
|
| D | vcvt-avx-int16-x16.c | 92 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local 120 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local
|
| D | vcvt-sse41-int16-x16.c | 92 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local 120 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local
|
| D | vcvt-neon-int16-x16.c | 89 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x16() local 116 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x16() local
|
| D | vcvt-sse2-int16-x16.c | 96 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16() local 126 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16() local
|
| D | vcvt-avx-int32-x16.c | 95 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local 123 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__avx_int32_x16() local
|
| D | vcvt-sse2-int32-x16.c | 99 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16() local 129 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse2_int32_x16() local
|
| D | vcvt-avx-int16-x24.c | 108 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() local 136 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() local
|
| D | vcvt-neon-int32-x16.c | 92 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local 119 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… in xnn_f16_f32_vcvt_ukernel__neon_int32_x16() local
|
| D | vcvt-sse41-int16-x24.c | 108 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() local 136 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() local
|
| D | vcvt-wasmsimd-int16-x16.c | 98 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local 131 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local
|
| D | vcvt-sse41-int32-x16.c | 95 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local 123 …const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32… in xnn_f16_f32_vcvt_ukernel__sse41_int32_x16() local
|
| /external/XNNPACK/src/f16-f32-vcvt/ |
| D | neon-int32.c.in | 78 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… variable 105 …const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo,… variable
|
| D | wasmsimd-int32.c.in | 85 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset)… variable 116 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset)… variable
|
| D | neon-int16.c.in | 82 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); variable 109 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); variable
|
| D | wasmsimd-int16.c.in | 92 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … variable 125 …const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0, 8, 1, 9, … variable
|