| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-wasmrelaxedsimd-fma.c | 344 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
346 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
527 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
528 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
649 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
650 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
|
| D | up8x25-wasmrelaxedsimd-fma.c | 342 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
344 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
522 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
523 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
643 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
644 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
|
| D | up8x25-minmax-wasmrelaxedsimd-fma-acc2.c | 344 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
346 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
530 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
531 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
654 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
655 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
|
| D | up4x25-minmax-neonfma.c | 246 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
247 vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
356 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
357 vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
|
| D | up4x25-minmax-neon-acc2.c | 246 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
247 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
358 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
359 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
|
| D | up4x25-minmax-neonfma-acc2.c | 246 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
247 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
358 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
359 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
|
| D | up4x25-minmax-wasmrelaxedsimd-fma.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
286 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
408 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
409 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
|
| D | up4x25-wasmrelaxedsimd-fma.c | 283 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
284 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
404 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
405 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
|
| D | up4x25-minmax-wasmrelaxedsimd.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
286 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
408 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
409 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
|
| D | up4x25-minmax-neon.c | 246 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
247 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
356 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
357 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
|
| D | up4x25-minmax-wasmrelaxedsimd-fma-acc2.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
286 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
410 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
411 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
|
| D | up4x25-minmax-sse-acc2.c | 285 const __m128 vk19x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
409 const __m128 vk19x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
410 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
|
| D | up4x25-minmax-sse.c | 285 const __m128 vk19x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
286 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
407 const __m128 vk19x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
408 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
|
| D | up4x25-minmax-wasmsimd-arm-acc2.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
410 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
411 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
|
| D | up4x25-minmax-wasmsimd-x86-acc2.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
410 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
411 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
|
| D | up4x25-minmax-wasmsimd-arm.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
286 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
408 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
409 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
|
| D | up4x25-minmax-wasmrelaxedsimd-acc2.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
410 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
411 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
|
| D | up4x25-minmax-wasmsimd-x86.c | 285 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
286 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
408 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
409 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
|
| D | up4x25-wasmsimd.c | 283 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
284 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
404 const v128_t vk19x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
405 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
|
| D | up8x25-minmax-sse-acc2.c | 344 const __m128 vk19x0123 = _mm_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
346 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
529 const __m128 vk19x0123 = _mm_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
530 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
653 const __m128 vk19x0123 = _mm_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
654 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
|
| D | up8x25-minmax-neonfma-acc2.c | 305 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
307 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
438 const float32x4_t vk19x0123 = vld1q_f32(w + 156); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
439 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
550 const float32x4_t vk19x0123 = vld1q_f32(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
551 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
|
| D | up8x25-minmax-neonfma.c | 305 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
307 vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
435 const float32x4_t vk19x0123 = vld1q_f32(w + 156); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
436 vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
545 const float32x4_t vk19x0123 = vld1q_f32(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
546 vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
|
| D | up8x25-minmax-neon.c | 305 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
307 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
435 const float32x4_t vk19x0123 = vld1q_f32(w + 156); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
436 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
545 const float32x4_t vk19x0123 = vld1q_f32(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
546 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
|
| D | up8x25-minmax-neon-acc2.c | 305 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
307 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
438 const float32x4_t vk19x0123 = vld1q_f32(w + 156); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
439 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
550 const float32x4_t vk19x0123 = vld1q_f32(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
551 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
|
| D | up8x25-minmax-wasmsimd-arm-acc2.c | 344 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
346 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
530 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
531 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
654 const v128_t vk19x0123 = wasm_v128_load(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
655 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
|