| /external/XNNPACK/src/f32-ibilinear/gen/ |
| D | neonfma-c8.c | 61 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
66 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
96 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
99 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
124 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
127 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
|
| D | neonfma-c4.c | 55 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c4() local
58 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c4() local
82 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c4() local
85 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c4() local
|
| D | neon-c8.c | 56 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local
79 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local
98 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local
|
| D | sse-c8.c | 63 const __m128 vt0123 = _mm_add_ps(vtl0123, _mm_mul_ps(vtd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local
91 const __m128 vt0123 = _mm_add_ps(vtl0123, _mm_mul_ps(vtd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local
110 const __m128 vt0123 = _mm_add_ps(vtl0123, _mm_mul_ps(vtd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local
|
| D | neon-c4.c | 50 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c4() local
68 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c4() local
|
| D | sse-c4.c | 57 const __m128 vt0123 = _mm_add_ps(vtl0123, _mm_mul_ps(vtd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c4() local
76 const __m128 vt0123 = _mm_add_ps(vtl0123, _mm_mul_ps(vtd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c4() local
|
| /external/XNNPACK/src/s8-ibilinear/gen/ |
| D | neon-c16.c | 68 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
78 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
122 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
128 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
162 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
168 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c16() local
|
| D | neon-c8.c | 59 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c8() local
65 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c8() local
99 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c8() local
105 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_s8_ibilinear_ukernel__neon_c8() local
|
| D | sse2-c16.c | 75 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse2_c16() local
132 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse2_c16() local
170 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse2_c16() local
|
| D | sse41-c16.c | 66 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse41_c16() local
114 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse41_c16() local
145 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse41_c16() local
|
| D | wasmsimd-dot16x2-c16.c | 64 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
111 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
141 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
|
| D | wasmsimd-mul32-c16.c | 67 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_s8_ibilinear_ukernel__wasmsimd_mul32_c16() local
114 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_s8_ibilinear_ukernel__wasmsimd_mul32_c16() local
146 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_s8_ibilinear_ukernel__wasmsimd_mul32_c16() local
|
| D | sse41-c8.c | 62 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse41_c8() local
93 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse41_c8() local
|
| D | wasmsimd-dot16x2-c8.c | 60 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c8() local
90 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c8() local
|
| D | sse2-c8.c | 67 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse2_c8() local
105 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_s8_ibilinear_ukernel__sse2_c8() local
|
| /external/XNNPACK/src/u8-ibilinear/gen/ |
| D | neon-c16.c | 68 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
78 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
122 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
128 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
162 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
168 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c16() local
|
| D | neon-c8.c | 59 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c8() local
65 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c8() local
99 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c8() local
105 …const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01… in xnn_u8_ibilinear_ukernel__neon_c8() local
|
| D | wasmsimd-dot16x2-c16.c | 64 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
111 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
141 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c16() local
|
| D | sse41-c16.c | 66 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse41_c16() local
114 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse41_c16() local
145 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse41_c16() local
|
| D | sse2-c16.c | 76 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse2_c16() local
134 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse2_c16() local
173 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse2_c16() local
|
| D | wasmsimd-mul32-c16.c | 67 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_u8_ibilinear_ukernel__wasmsimd_mul32_c16() local
114 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_u8_ibilinear_ukernel__wasmsimd_mul32_c16() local
146 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_u8_ibilinear_ukernel__wasmsimd_mul32_c16() local
|
| D | wasmsimd-dot16x2-c8.c | 60 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c8() local
90 …const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9… in xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c8() local
|
| D | sse41-c8.c | 62 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse41_c8() local
93 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse41_c8() local
|
| D | sse2-c8.c | 68 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse2_c8() local
107 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); in xnn_u8_ibilinear_ukernel__sse2_c8() local
|
| D | wasmsimd-mul32-c8.c | 59 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_u8_ibilinear_ukernel__wasmsimd_mul32_c8() local
91 …const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11),… in xnn_u8_ibilinear_ukernel__wasmsimd_mul32_c8() local
|