Home
last modified time | relevance | path

Searched refs:vi21x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c294 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
415 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c294 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
417 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c294 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
414 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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.c294 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
416 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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.c292 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
411 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c253 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
363 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c294 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
415 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c253 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
365 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c294 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
417 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c253 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
365 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c253 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
363 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c358 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
535 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
655 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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.c317 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
442 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
444 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
552 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c317 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
442 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
444 vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
552 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c356 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
531 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
650 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c317 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
445 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; 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()
557 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c358 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
536 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
656 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c317 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 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()
445 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; 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()
557 const float32x4_t vi21x0123 = vld1q_f32(i21); 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.c358 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
536 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
656 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c358 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
538 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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()
660 const __m128 vi21x0123 = _mm_loadu_ps(i21); 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.c358 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
539 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
661 const v128_t vi21x0123 = wasm_v128_load(i21); 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.c358 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
539 const v128_t vi21x0123 = wasm_v128_load(i21); 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()
661 const v128_t vi21x0123 = wasm_v128_load(i21); 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()