/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-wasmsimd.c | 163 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 232 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 279 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up4x9-wasmsimd.c | 137 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 184 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up8x9-minmax-sse.c | 165 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 236 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 284 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|
D | up8x9-minmax-neon.c | 148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 199 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 245 const float32x4_t vk8x0123 = vld1q_f32(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
|
D | up8x9-minmax-sse-acc2.c | 165 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 239 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 289 const __m128 vk8x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
|
D | up8x9-minmax-neonfma.c | 148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 199 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 245 const float32x4_t vk8x0123 = vld1q_f32(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 165 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 240 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 290 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 165 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 240 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 290 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|
D | up8x9-minmax-wasmsimd-x86.c | 165 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 237 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 285 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
|
D | up8x9-minmax-neonfma-acc2.c | 148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 202 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 250 const float32x4_t vk8x0123 = vld1q_f32(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-minmax-wasmsimd-arm.c | 165 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 237 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 285 const v128_t vk8x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
|
D | up8x9-minmax-neon-acc2.c | 148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 202 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 250 const float32x4_t vk8x0123 = vld1q_f32(w + 72); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|
D | up4x9-minmax-wasmsimd-arm.c | 139 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local 188 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
|
D | up4x9-minmax-wasmsimd-x86-acc2.c | 139 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local 190 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
|
D | up4x9-minmax-sse.c | 139 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local 187 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
|
D | up4x9-minmax-neon-acc2.c | 122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local 170 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
|
D | up4x9-minmax-neonfma.c | 122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local 168 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
|
D | up4x9-minmax-sse-acc2.c | 139 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local 189 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
|
D | up4x9-minmax-wasmsimd-x86.c | 139 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local 188 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
|
D | up4x9-minmax-neonfma-acc2.c | 122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local 170 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-minmax-neon.c | 122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local 168 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
|
D | up4x9-minmax-wasmsimd-arm-acc2.c | 139 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local 190 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
|
D | up4x25-wasmsimd.c | 217 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 360 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 219 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 366 const v128_t vk8x0123 = wasm_v128_load(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 219 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 363 const __m128 vk8x0123 = _mm_load_ps(w + 36); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|