| /external/XNNPACK/src/f32-f16-vcvt/gen/ |
| D | vcvt-scalar-fabsf-x4.c | 67 uint32_t vbias3 = vnonsignw3 + vexp_bias; in xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4() local
80 vbias3 &= vexpw_max; in xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4()
85 vbias3 = math_max_u32(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4()
90 vf3 += uint32_as_float(vbias3); in xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x4()
|
| D | vcvt-scalar-bitcast-x4.c | 63 uint32_t vbias3 = vnonsignw3 + vexp_bias; in xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4() local
72 vbias3 &= vexpw_max; in xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4()
81 vbias3 = math_max_u32(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4()
86 vf3 += uint32_as_float(vbias3); in xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4()
|
| D | vcvt-neon-x16.c | 54 uint32x4_t vbias3 = vaddq_u32(vreinterpretq_u32_f32(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__neon_x16() local
68 vbias3 = vandq_u32(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__neon_x16()
79 vbias3 = vmaxq_u32(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__neon_x16()
84 vf3 = vaddq_f32(vf3, vreinterpretq_f32_u32(vbias3)); in xnn_f32_f16_vcvt_ukernel__neon_x16()
|
| D | vcvt-wasmsimd-x16.c | 59 v128_t vbias3 = wasm_i32x4_add(vabsx3, vexp_bias); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x16() local
74 vbias3 = wasm_v128_and(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x16()
90 vbias3 = wasm_i16x8_max(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x16()
95 vf3 = wasm_f32x4_add(vf3, vbias3); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x16()
|
| D | vcvt-neon-x24.c | 58 uint32x4_t vbias3 = vaddq_u32(vreinterpretq_u32_f32(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__neon_x24() local
78 vbias3 = vandq_u32(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__neon_x24()
94 vbias3 = vmaxq_u32(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__neon_x24()
101 vf3 = vaddq_f32(vf3, vreinterpretq_f32_u32(vbias3)); in xnn_f32_f16_vcvt_ukernel__neon_x24()
|
| D | vcvt-sse41-x16.c | 61 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse41_x16() local
76 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse41_x16()
92 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse41_x16()
98 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse41_x16()
|
| D | vcvt-avx-x16.c | 61 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__avx_x16() local
76 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__avx_x16()
92 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__avx_x16()
98 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__avx_x16()
|
| D | vcvt-sse2-x16.c | 61 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse2_x16() local
76 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
92 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
100 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
|
| D | vcvt-wasmsimd-x24.c | 65 v128_t vbias3 = wasm_i32x4_add(vabsx3, vexp_bias); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x24() local
86 vbias3 = wasm_v128_and(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x24()
108 vbias3 = wasm_i16x8_max(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x24()
115 vf3 = wasm_f32x4_add(vf3, vbias3); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x24()
|
| D | vcvt-neon-x32.c | 62 uint32x4_t vbias3 = vaddq_u32(vreinterpretq_u32_f32(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__neon_x32() local
88 vbias3 = vandq_u32(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__neon_x32()
109 vbias3 = vmaxq_u32(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__neon_x32()
118 vf3 = vaddq_f32(vf3, vreinterpretq_f32_u32(vbias3)); in xnn_f32_f16_vcvt_ukernel__neon_x32()
|
| D | vcvt-sse2-x24.c | 67 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse2_x24() local
88 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse2_x24()
110 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse2_x24()
121 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse2_x24()
|
| D | vcvt-sse41-x24.c | 67 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse41_x24() local
88 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse41_x24()
110 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse41_x24()
118 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse41_x24()
|
| D | vcvt-avx-x24.c | 67 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__avx_x24() local
88 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__avx_x24()
110 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__avx_x24()
118 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__avx_x24()
|
| D | vcvt-wasmsimd-x32.c | 71 v128_t vbias3 = wasm_i32x4_add(vabsx3, vexp_bias); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x32() local
98 vbias3 = wasm_v128_and(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x32()
126 vbias3 = wasm_i16x8_max(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x32()
135 vf3 = wasm_f32x4_add(vf3, vbias3); in xnn_f32_f16_vcvt_ukernel__wasmsimd_x32()
|
| D | vcvt-avx-x32.c | 73 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__avx_x32() local
100 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__avx_x32()
128 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__avx_x32()
138 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__avx_x32()
|
| D | vcvt-sse2-x32.c | 73 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse2_x32() local
100 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse2_x32()
128 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse2_x32()
142 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse2_x32()
|
| D | vcvt-sse41-x32.c | 73 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse41_x32() local
100 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse41_x32()
128 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse41_x32()
138 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse41_x32()
|
| /external/XNNPACK/src/f32-vmulcaddc/gen/ |
| D | c4-minmax-scalar-2x.c | 68 const float vbias3 = w[7]; in xnn_f32_vmulcaddc_minmax_ukernel_c4__scalar_2x() local
73 vacc0x3 = vacc0x3 * vscale3 + vbias3; in xnn_f32_vmulcaddc_minmax_ukernel_c4__scalar_2x()
77 vacc1x3 = vacc1x3 * vscale3 + vbias3; in xnn_f32_vmulcaddc_minmax_ukernel_c4__scalar_2x()
|
| D | c4-minmax-wasm-2x.c | 68 const float vbias3 = w[7]; in xnn_f32_vmulcaddc_minmax_ukernel_c4__wasm_2x() local
73 vacc0x3 = vacc0x3 * vscale3 + vbias3; in xnn_f32_vmulcaddc_minmax_ukernel_c4__wasm_2x()
77 vacc1x3 = vacc1x3 * vscale3 + vbias3; in xnn_f32_vmulcaddc_minmax_ukernel_c4__wasm_2x()
|
| /external/XNNPACK/src/amalgam/ |
| D | avx.c | 1282 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__avx_x24() local
1303 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__avx_x24()
1325 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__avx_x24()
1333 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__avx_x24()
|
| D | sse2.c | 1037 __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); in xnn_f32_f16_vcvt_ukernel__sse2_x16() local
1052 vbias3 = _mm_and_si128(vbias3, vexpw_max); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
1068 vbias3 = _mm_max_epi16(vbias3, vbias_min); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
1076 vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); in xnn_f32_f16_vcvt_ukernel__sse2_x16()
|