Home
last modified time | relevance | path

Searched refs:vk23x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
310 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
424 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
425 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
426 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
427 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c309 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
310 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
423 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
424 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c309 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
310 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
425 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
426 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c307 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
308 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
420 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
421 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
263 vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
372 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
373 vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
310 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
424 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
425 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
263 vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
374 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
375 vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
426 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
427 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
263 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
374 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
375 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
263 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
372 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
373 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c380 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
382 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
550 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
551 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
664 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
665 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
335 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
451 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
452 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
561 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
562 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
335 vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
451 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
452 vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
561 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
562 vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c378 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
380 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
546 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
547 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
659 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
660 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
335 vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
454 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
455 vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
566 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
567 vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
382 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
551 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
552 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
665 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
666 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
335 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
454 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
455 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
566 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
567 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
382 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
551 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
552 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
665 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
666 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c380 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
382 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
553 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
554 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
669 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
670 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
382 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
554 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
555 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
670 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
671 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
382 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
554 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
555 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
670 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
671 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()