| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x9-minmax-sse.c | 130 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
132 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
|
| D | up8x9-wasmsimd.c | 128 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
130 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
|
| D | up8x9-minmax-neon.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
123 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
|
| D | up8x9-minmax-sse-acc2.c | 130 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
132 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
|
| D | up8x9-minmax-neonfma.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
123 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
|
| D | up8x9-minmax-wasmsimd-x86.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
|
| D | up8x9-minmax-neonfma-acc2.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
123 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-minmax-wasmsimd-arm.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
|
| D | up8x9-minmax-neon-acc2.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
123 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
|
| D | up8x25-minmax-neon-acc2.c | 201 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
203 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
|
| D | up8x25-minmax-sse.c | 210 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
212 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
|
| D | up8x25-minmax-wasmsimd-arm.c | 210 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
212 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
|
| D | up8x25-minmax-neon.c | 201 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
203 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
|
| D | up8x25-minmax-wasmsimd-arm-acc2.c | 210 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
212 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
|
| D | up8x25-minmax-wasmsimd-x86-acc2.c | 210 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
212 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
|
| D | up8x25-minmax-neonfma.c | 201 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
203 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
|
| D | up8x25-minmax-sse-acc2.c | 210 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
212 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
|
| D | up8x25-wasmsimd.c | 208 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
210 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
|
| D | up8x25-minmax-wasmsimd-x86.c | 210 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
212 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
|
| D | up8x25-minmax-neonfma-acc2.c | 201 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
203 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
|