/external/XNNPACK/src/f32-argmaxpool/ |
D | 9p8x-neon-c4.c | 154 const uint32x4_t vidx2 = vaddq_u32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local 257 const uint32x4_t vidx2 = vaddq_u32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local 312 const uint32x4_t vidx2 = vaddq_u32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
|
D | 9p8x-wasmsimd-c4.c | 173 const v128_t vidx2 = wasm_i32x4_add(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local 288 const v128_t vidx2 = wasm_i32x4_add(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local 345 const v128_t vidx2 = wasm_i32x4_add(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
|
D | 9p8x-sse2-c4.c | 173 const __m128i vidx2 = _mm_add_epi32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local 288 const __m128i vidx2 = _mm_add_epi32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local 345 const __m128i vidx2 = _mm_add_epi32(vidx1, v1); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
|
/external/XNNPACK/src/math/ |
D | expm1minus-sse2-rr2-lut16-p3.c | 82 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx, 4); in xnn_math_f32_expm1minus__sse2_rr2_lut16_p3() local
|
D | sigmoid-sse2-rr2-lut64-p2-div.c | 87 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx, 4); in xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_div() local
|
D | exp-sse2-rr2-lut64-p2.c | 80 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx, 4); in xnn_math_f32_exp__sse2_rr2_lut64_p2() local
|
D | sigmoid-sse2-rr2-lut64-p2-nr2.c | 88 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx, 4); in xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_nr2() local
|
D | sigmoid-sse2-rr2-lut64-p2-nr1.c | 88 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx, 4); in xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_nr1() local
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-wasm-rr2-lut16-p3-x3.c | 64 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x3() local
|
D | velu-scalar-rr2-lut16-p3-x3.c | 64 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x3() local
|
D | velu-scalar-rr2-lut16-p3-x4.c | 67 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x4() local
|
D | velu-wasm-rr2-lut16-p3-x4.c | 67 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x4() local
|
D | velu-avx2-rr1-lut16-p3-gather-x24.c | 60 const __m256i vidx2 = _mm256_and_si256(_mm256_castps_si256(vn2), vindex_mask); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x24() local
|
D | velu-scalar-rr2-lut16-p3-x5.c | 70 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x5() local
|
D | velu-wasm-rr2-lut16-p3-x5.c | 70 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x5() local
|
D | velu-avx2-rr1-lut16-p3-gather-x32.c | 63 const __m256i vidx2 = _mm256_and_si256(_mm256_castps_si256(vn2), vindex_mask); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x32() local
|
D | velu-scalar-rr2-lut16-p3-x6.c | 73 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6() local
|
D | velu-wasm-rr2-lut16-p3-x6.c | 73 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6() local
|
D | velu-avx2-rr1-lut16-p3-gather-x40.c | 66 const __m256i vidx2 = _mm256_and_si256(_mm256_castps_si256(vn2), vindex_mask); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x40() local
|
D | velu-avx2-rr1-lut16-p3-gather-x48.c | 69 const __m256i vidx2 = _mm256_and_si256(_mm256_castps_si256(vn2), vindex_mask); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | scalar-lut64-p2-x4.c | 87 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4() local
|
D | scalar-lut64-p2-x4-acc2.c | 88 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc2() local
|
D | scalar-lut64-p2-x4-acc4.c | 90 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc4() local
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | scalar-lut2048-p1-div-x4.c | 65 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4() local
|
D | scalar-lut64-p2-div-x4.c | 65 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4() local
|