/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-wasmsimd.c | 123 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 205 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 262 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up4x9-wasmsimd.c | 110 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 167 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up8x9-minmax-sse.c | 125 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 209 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 267 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|
D | up8x9-minmax-neon.c | 118 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 182 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 228 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
|
D | up8x9-minmax-sse-acc2.c | 125 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 212 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 272 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
|
D | up8x9-minmax-neonfma.c | 118 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 182 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 228 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 125 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 213 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 273 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 125 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 213 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 273 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|
D | up8x9-minmax-wasmsimd-x86.c | 125 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 210 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 268 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
|
D | up8x9-minmax-neonfma-acc2.c | 118 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 185 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 233 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-minmax-wasmsimd-arm.c | 125 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 210 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 268 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
|
D | up8x9-minmax-neon-acc2.c | 118 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 185 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 233 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|
D | up4x9-minmax-wasmsimd-arm.c | 112 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local 171 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
|
D | up4x9-minmax-wasmsimd-x86-acc2.c | 112 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local 173 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
|
D | up4x9-minmax-sse.c | 112 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local 170 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
|
D | up4x9-minmax-neon-acc2.c | 105 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local 153 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
|
D | up4x9-minmax-neonfma.c | 105 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local 151 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
|
D | up4x9-minmax-sse-acc2.c | 112 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local 172 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
|
D | up4x9-minmax-wasmsimd-x86.c | 112 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local 171 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
|
D | up4x9-minmax-neonfma-acc2.c | 105 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local 153 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-minmax-neon.c | 105 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local 151 const float32x4_t vi4x0123 = vld1q_f32(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
|
D | up4x9-minmax-wasmsimd-arm-acc2.c | 112 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local 173 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
|
D | up4x25-wasmsimd.c | 190 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 343 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 192 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 349 const v128_t vi4x0123 = wasm_v128_load(i4); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 192 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 346 const __m128 vi4x0123 = _mm_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|