Home
last modified time | relevance | path

Searched refs:vk14x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c255 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
390 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c255 const __m128 vk14x0123 = _mm_load_ps(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
256 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
387 const __m128 vk14x0123 = _mm_load_ps(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
388 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c226 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
227 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
338 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
339 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c226 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
227 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
336 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
337 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c226 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
227 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
336 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
337 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c253 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
254 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
384 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
385 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c226 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
227 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
338 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
339 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c255 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
388 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
389 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c255 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
390 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c255 const __m128 vk14x0123 = _mm_load_ps(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
256 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
389 const __m128 vk14x0123 = _mm_load_ps(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
390 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c255 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
388 const v128_t vk14x0123 = wasm_v128_load(w + 60); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
389 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c270 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
272 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
418 const float32x4_t vk14x0123 = vld1q_f32(w + 116); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
419 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
530 const float32x4_t vk14x0123 = vld1q_f32(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
531 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c299 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
301 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
496 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
497 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
628 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
629 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c299 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
301 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
497 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
498 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
629 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
630 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c270 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
272 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
415 const float32x4_t vk14x0123 = vld1q_f32(w + 116); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
416 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
525 const float32x4_t vk14x0123 = vld1q_f32(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
526 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c299 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
301 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
500 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
501 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
634 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
635 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c299 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
301 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
500 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
501 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
634 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
635 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c270 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
272 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
415 const float32x4_t vk14x0123 = vld1q_f32(w + 116); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
416 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
525 const float32x4_t vk14x0123 = vld1q_f32(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
526 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c299 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
301 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
499 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
500 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
633 const __m128 vk14x0123 = _mm_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
634 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c297 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
299 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
492 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
493 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
623 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
624 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c299 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
301 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
497 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
498 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
629 const v128_t vk14x0123 = wasm_v128_load(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
630 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c270 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
272 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
418 const float32x4_t vk14x0123 = vld1q_f32(w + 116); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
419 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
530 const float32x4_t vk14x0123 = vld1q_f32(w + 120); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
531 vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()