| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 12x2-minmax-neonfma.c | 52 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 102 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 149 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 188 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
|
| D | 32x2-minmax-neonfma.c | 62 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 163 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 240 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 299 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 347 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 386 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
|
| D | 16x2-minmax-neonfma.c | 54 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 114 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 166 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 205 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm.c | 54 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 120 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 163 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 134 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 177 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
|
| D | 32x1-minmax-neon.c | 54 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 123 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 165 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
|
| D | 32x1-minmax-neonfma-pipelined.c | 42 float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 134 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 179 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined-x2.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 174 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 217 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local
|
| D | 32x1-minmax-wasmsimd-x86.c | 54 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 120 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 163 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
|
| D | 32x1-minmax-neonfma.c | 54 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 123 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 165 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 134 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 177 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 174 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 217 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
|
| D | 32x1-minmax-sse.c | 54 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 120 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 162 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
|
| D | 32x1-minmax-neon-pipelined.c | 42 float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 134 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 179 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
|
| D | 12x4-minmax-neonfma.c | 58 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 134 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 185 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 242 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local
|
| D | 32x4-minmax-neonfma.c | 78 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 245 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 330 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 423 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 475 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 532 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
|
| D | 16x4-minmax-neonfma.c | 62 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 156 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 212 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 269 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local
|
| D | 16x1-minmax-neon-pipelined.c | 42 float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local 103 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local
|
| D | 16x1-minmax-neon.c | 50 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local 96 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local
|
| D | 16x1-minmax-wasmsimd-arm-pipelined-x2.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2() local 128 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2() local
|
| D | 16x1-minmax-wasmsimd-x86-pipelined-x2.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined_x2() local 128 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined_x2() local
|
| D | 16x1-minmax-wasmsimd-x86-pipelined.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local 104 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local
|
| D | 16x1-minmax-neonfma.c | 50 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local 96 const float32x4_t vi4567 = vld1q_f32(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local
|
| D | 16x1-minmax-wasmsimd-x86.c | 50 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local 94 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local
|
| D | 16x1-minmax-wasmsimd-arm-pipelined.c | 42 v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined() local 104 const v128_t vi4567 = wasm_v128_load(input + 4); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined() local
|