| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 32x2-minmax-neonfma.c | 64 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 165 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 242 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 301 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm.c | 56 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 122 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined.c | 44 v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 136 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
|
| D | 32x1-minmax-neon.c | 56 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 125 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
|
| D | 32x1-minmax-neonfma-pipelined.c | 44 float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 136 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined-x2.c | 44 v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 176 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local
|
| D | 32x1-minmax-wasmsimd-x86.c | 56 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 122 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
|
| D | 32x1-minmax-neonfma.c | 56 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 125 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined.c | 44 v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 136 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 44 v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 176 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
|
| D | 32x1-minmax-sse.c | 56 const __m128 viCDEF = _mm_loadu_ps(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 122 const __m128 viCDEF = _mm_loadu_ps(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
|
| D | 32x1-minmax-neon-pipelined.c | 44 float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 136 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
|
| D | 32x4-minmax-neonfma.c | 80 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 247 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 332 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 425 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
|
| D | 16x2-minmax-neonfma.c | 56 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 116 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm-x2.c | 124 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2() local 190 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2() local
|
| D | 32x1-minmax-neon-x2.c | 127 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() local 196 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() local
|
| D | 32x1-minmax-neonfma-x2.c | 127 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2() local 196 const float32x4_t viCDEF = vld1q_f32(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2() local
|
| D | 32x1-minmax-wasmsimd-x86-x2.c | 124 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local 190 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | sse2-rr2-p5-x16.c | 46 const __m128 viCDEF = _mm_loadu_ps(input + 12); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x16() local
|
| D | sse2-rr2-p5-x16-acc2.c | 47 const __m128 viCDEF = _mm_loadu_ps(input + 12); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x16_acc2() local
|
| D | sse2-rr2-p5-x16-acc4.c | 49 const __m128 viCDEF = _mm_loadu_ps(input + 12); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x16_acc4() local
|
| D | wasmsimd-rr2-p5-x16.c | 46 const v128_t viCDEF = wasm_v128_load(input + 12); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x16() local
|
| D | neonfma-rr1-p5-x16.c | 44 const float32x4_t viCDEF = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16() local
|
| D | neonfma-rr1-p5-x16-acc4.c | 47 const float32x4_t viCDEF = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x16_acc4() local
|
| D | neon-rr2-p5-x16-acc2.c | 46 const float32x4_t viCDEF = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16_acc2() local
|