Home
last modified time | relevance | path

Searched refs:vi24x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c312 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
316 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
429 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
431 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c312 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
316 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
426 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
428 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c265 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
267 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
377 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
379 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c265 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
267 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
375 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
377 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c265 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
267 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
375 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
377 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c310 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
314 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
423 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
425 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c265 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
267 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
377 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
379 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c312 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
316 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
427 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
429 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c312 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
316 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
429 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
431 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c312 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
316 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
428 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
430 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c312 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
316 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
427 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
429 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c338 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
342 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
457 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
459 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
569 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
571 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c385 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
391 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
553 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
557 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
667 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
669 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c385 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
554 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
558 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
668 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
670 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c338 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
342 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
454 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
456 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
564 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
566 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c385 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
557 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
561 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
673 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
675 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c385 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
557 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
561 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
673 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
675 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c338 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
342 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
454 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
456 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
564 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
566 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c385 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
391 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
556 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
560 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
672 const __m128 vi24x0123 = _mm_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
674 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c383 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
389 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
549 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
553 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
662 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
664 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c385 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
554 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
558 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
668 const v128_t vi24x0123 = wasm_v128_load(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
670 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c338 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
342 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
457 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
459 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
569 const float32x4_t vi24x0123 = vld1q_f32(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
571 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()