/external/XNNPACK/src/f16-f32-vcvt/gen/ |
D | vcvt-wasmsimd-int16-x16.c | 50 const v128_t vprenorm2 = wasm_i16x8_shl(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() local 55 …const v128_t vnorm2 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 0, 8, 1, 9, 2, 10,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16() 56 …const v128_t vnorm3 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 4, 12, 5, 13, 6, 14,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16()
|
D | vcvt-sse41-int16-x16.c | 50 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local 55 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() 56 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16()
|
D | vcvt-avx-int16-x16.c | 50 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local 55 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() 56 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16()
|
D | vcvt-wasmsimd-int16-x24.c | 53 const v128_t vprenorm2 = wasm_i16x8_shl(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24() local 60 …const v128_t vnorm2 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 0, 8, 1, 9, 2, 10,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24() 61 …const v128_t vnorm3 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 4, 12, 5, 13, 6, 14,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x24()
|
D | vcvt-sse41-int16-x24.c | 53 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() local 60 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24() 61 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x24()
|
D | vcvt-sse2-int16-x16.c | 50 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16() local 55 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16() 56 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x16()
|
D | vcvt-avx-int16-x24.c | 53 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() local 60 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24() 61 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x24()
|
D | vcvt-sse2-int16-x24.c | 53 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24() local 60 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24() 61 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x24()
|
D | vcvt-neon-int16-x24.c | 51 …const uint16x8x2_t vprenorm2 = vzipq_u16(vshlq_n_u16(vnonsign2, 13), vsraq_n_u16(vexp_offset, vnon… in xnn_f16_f32_vcvt_ukernel__neon_int16_x24() local 57 const float32x4_t vnorm4 = vmulq_f32(vreinterpretq_f32_u16(vprenorm2.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x24() 58 const float32x4_t vnorm5 = vmulq_f32(vreinterpretq_f32_u16(vprenorm2.val[1]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x24()
|
D | vcvt-wasmsimd-int16-x32.c | 56 const v128_t vprenorm2 = wasm_i16x8_shl(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32() local 65 …const v128_t vnorm2 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 0, 8, 1, 9, 2, 10,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32() 66 …const v128_t vnorm3 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 4, 12, 5, 13, 6, 14,… in xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x32()
|
D | vcvt-sse41-int16-x32.c | 56 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32() local 65 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32() 66 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x32()
|
D | vcvt-avx-int16-x32.c | 56 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__avx_int16_x32() local 65 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x32() 66 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x32()
|
D | vcvt-neon-int16-x32.c | 54 …const uint16x8x2_t vprenorm2 = vzipq_u16(vshlq_n_u16(vnonsign2, 13), vsraq_n_u16(vexp_offset, vnon… in xnn_f16_f32_vcvt_ukernel__neon_int16_x32() local 61 const float32x4_t vnorm4 = vmulq_f32(vreinterpretq_f32_u16(vprenorm2.val[0]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x32() 62 const float32x4_t vnorm5 = vmulq_f32(vreinterpretq_f32_u16(vprenorm2.val[1]), vexp_scale); in xnn_f16_f32_vcvt_ukernel__neon_int16_x32()
|
D | vcvt-sse2-int16-x32.c | 56 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() local 65 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() 66 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32()
|
/external/XNNPACK/src/amalgam/ |
D | sse41.c | 60 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() local 65 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16() 66 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse41_int16_x16()
|
D | avx.c | 59 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() local 64 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16() 65 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__avx_int16_x16()
|
D | sse2.c | 75 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() local 84 …norm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32() 85 …norm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), ve… in xnn_f16_f32_vcvt_ukernel__sse2_int16_x32()
|