Home
last modified time | relevance | path

Searched refs:vk21x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c297 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
298 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
416 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
417 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c297 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
298 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
418 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
419 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c297 const __m128 vk21x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
298 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
415 const __m128 vk21x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
416 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c297 const __m128 vk21x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
298 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
417 const __m128 vk21x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
418 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c295 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
296 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
412 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
413 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c254 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
255 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
364 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
365 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c297 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
298 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
416 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
417 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c254 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
255 vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
366 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
367 vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c297 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
298 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
418 const v128_t vk21x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
419 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c254 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
255 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
366 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
367 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c254 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
255 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
364 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
365 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c362 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
364 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
538 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
539 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
656 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
657 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c319 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
321 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
443 const float32x4_t vk21x0123 = vld1q_f32(w + 172); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
444 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
553 const float32x4_t vk21x0123 = vld1q_f32(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
554 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c319 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
321 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
443 const float32x4_t vk21x0123 = vld1q_f32(w + 172); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
444 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
553 const float32x4_t vk21x0123 = vld1q_f32(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
554 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c360 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
362 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
534 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
535 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
651 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
652 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c319 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
321 vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
446 const float32x4_t vk21x0123 = vld1q_f32(w + 172); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
447 vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
558 const float32x4_t vk21x0123 = vld1q_f32(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
559 vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c362 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
364 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
539 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
540 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
657 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
658 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c319 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
321 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
446 const float32x4_t vk21x0123 = vld1q_f32(w + 172); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
447 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
558 const float32x4_t vk21x0123 = vld1q_f32(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
559 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c362 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
364 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
539 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
540 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
657 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
658 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c362 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
364 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
541 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
542 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
661 const __m128 vk21x0123 = _mm_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
662 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c362 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
364 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
542 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
543 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
662 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
663 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c362 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
364 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
542 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
543 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
662 const v128_t vk21x0123 = wasm_v128_load(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
663 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()