/external/XNNPACK/src/f32-velu/gen/ |
D | velu-neonfma-rr1-lut16-p3-x12.c | 62 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12() local 91 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12()
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x12.c | 64 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12() local 93 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x12.c | 64 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() local 93 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
|
D | velu-neon-rr2-lut16-p3-x12.c | 63 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12() local 92 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12()
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x16.c | 67 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x16() local 105 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x16()
|
D | velu-neon-rr2-lut16-p3-x16.c | 66 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x16() local 104 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x16()
|
D | velu-neonfma-rr1-lut16-p3-x16.c | 65 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16() local 103 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x16.c | 67 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() local 105 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
|
D | velu-sse41-rr2-lut16-p3-x12.c | 64 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local 123 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x20.c | 70 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() local 117 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20()
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x24.c | 73 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() local 129 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24()
|
D | velu-sse2-rr2-lut16-p3-x12.c | 64 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local 135 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
|
D | velu-neon-rr2-lut16-p3-x20.c | 69 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20() local 116 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20()
|
D | velu-neonfma-rr1-lut16-p3-x20.c | 68 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20() local 115 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x20.c | 70 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() local 117 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
|
D | velu-neonfma-rr1-lut16-p3-x24.c | 71 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() local 127 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
|
D | velu-sse41-rr2-lut16-p3-x16.c | 67 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 144 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x24.c | 73 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() local 129 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
|
D | velu-neon-rr2-lut16-p3-x24.c | 72 const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() local 128 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
|
D | velu-sse2-rr2-lut16-p3-x16.c | 67 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 160 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16()
|
D | velu-sse41-rr2-lut16-p3-x20.c | 70 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 165 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
|
D | velu-sse2-rr2-lut16-p3-x20.c | 70 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 185 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20()
|
D | velu-sse41-rr2-lut16-p3-x24.c | 73 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 186 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
|
D | velu-sse2-rr2-lut16-p3-x24.c | 73 const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local 210 __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24()
|