| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 135 v128_t vi8x4567 = wasm_v128_load(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
183 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
193 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
241 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
242 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
282 const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
330 const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
331 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
437 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
447 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-5x4.c | 106 float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
154 vo4p0 = vfmaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
164 const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
212 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
213 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
253 const float32x4_t vi8x5678 = vextq_f32(vi8x4567, vi8x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
301 const float32x4_t vi8x6789 = vextq_f32(vi8x4567, vi8x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
302 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
409 vo4p0 = vfmaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
419 const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
[all …]
|
| D | 5x5p2-minmax-neon-5x4.c | 106 float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
154 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
164 const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
212 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
213 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
253 const float32x4_t vi8x5678 = vextq_f32(vi8x4567, vi8x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
301 const float32x4_t vi8x6789 = vextq_f32(vi8x4567, vi8x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
302 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
409 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
419 const float32x4_t vi8x3456 = vextq_f32(vi8x0123, vi8x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 135 v128_t vi8x4567 = wasm_v128_load(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
183 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
193 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
241 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
242 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
282 const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
330 const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
331 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
437 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
447 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 109 v128_t vi8x4567 = wasm_v128_load(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
157 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
167 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
215 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
216 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
256 const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
304 const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
305 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
411 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
421 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 109 v128_t vi8x4567 = wasm_v128_load(i8); i8 += 4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
157 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
167 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
215 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
216 vi8x0123 = vi8x4567; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
256 const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
304 const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
305 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
411 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
421 const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-sse-5x4.c | 133 __m128 vi8x4567 = _mm_loadu_ps(i8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
162 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
172 const __m128 vi8x7456 = _mm_shuffle_ps(vi8x4567, vi8x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
264 const __m128 vi8x8567 = _mm_move_ss(vi8x4567, vi8x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
265 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
415 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
425 const __m128 vi8x7456 = _mm_shuffle_ps(vi8x4567, vi8x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
517 const __m128 vi8x8567 = _mm_move_ss(vi8x4567, vi8x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
518 vi8x4567 = vi8x89AB; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
655 vi8x4567 = _mm_and_ps(vi8x4567, vmask); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x9-minmax-sse.c | 162 const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
168 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
|
| D | up8x9-wasmsimd.c | 160 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
166 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
|
| D | up8x9-minmax-neon.c | 147 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
151 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
|
| D | up8x9-minmax-sse-acc2.c | 162 const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
168 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
|
| D | up8x9-minmax-neonfma.c | 147 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
151 vacc4567p0 = vfmaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 162 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
168 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
|
| D | up8x9-minmax-wasmsimd-x86.c | 162 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
168 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
|
| D | up8x9-minmax-neonfma-acc2.c | 147 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
151 vacc4567p0 = vfmaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-minmax-wasmsimd-arm.c | 162 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
168 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
|
| D | up8x9-minmax-neon-acc2.c | 147 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
151 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 162 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
168 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
|
| D | up8x25-minmax-neon-acc2.c | 227 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
231 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
|
| D | up8x25-minmax-sse.c | 242 const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
248 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
|
| D | up8x25-minmax-wasmsimd-arm.c | 242 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
248 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
|
| D | up8x25-minmax-neon.c | 227 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
231 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
|
| D | up8x25-minmax-wasmsimd-arm-acc2.c | 242 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
248 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
|
| D | up8x25-minmax-wasmsimd-x86-acc2.c | 242 const v128_t vi8x4567 = wasm_v128_load(i8 + 4); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
248 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
|
| D | up8x25-minmax-neonfma.c | 227 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
231 vacc4567p0 = vfmaq_f32(vacc4567p0, vi8x4567, vk8x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
|