Searched refs:vi20x4567 (Results 1 – 11 of 11) sorted by relevance
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-neon-acc2.c | 311 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
315 vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
|
| D | up8x25-minmax-sse.c | 350 const __m128 vi20x4567 = _mm_loadu_ps(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
356 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
|
| D | up8x25-minmax-wasmsimd-arm.c | 350 const v128_t vi20x4567 = wasm_v128_load(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
356 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
|
| D | up8x25-minmax-neon.c | 311 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
315 vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
|
| D | up8x25-minmax-wasmsimd-arm-acc2.c | 350 const v128_t vi20x4567 = wasm_v128_load(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
356 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
|
| D | up8x25-minmax-wasmsimd-x86-acc2.c | 350 const v128_t vi20x4567 = wasm_v128_load(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
356 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
|
| D | up8x25-minmax-neonfma.c | 311 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
315 vacc4567p0 = vfmaq_f32(vacc4567p0, vi20x4567, vk20x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
|
| D | up8x25-minmax-sse-acc2.c | 350 const __m128 vi20x4567 = _mm_loadu_ps(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
356 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
|
| D | up8x25-wasmsimd.c | 348 const v128_t vi20x4567 = wasm_v128_load(i20 + 4); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
354 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
|
| D | up8x25-minmax-wasmsimd-x86.c | 350 const v128_t vi20x4567 = wasm_v128_load(i20 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
356 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
|
| D | up8x25-minmax-neonfma-acc2.c | 311 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
315 vacc4567p0 = vfmaq_f32(vacc4567p0, vi20x4567, vk20x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
|