Home
last modified time | relevance | path

Searched refs:vk20x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
292 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
413 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
292 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
415 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c291 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
292 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
411 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
412 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c291 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
292 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
413 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
414 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c289 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
290 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
408 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
409 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
251 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
360 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
361 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
292 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
413 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
251 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
362 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
363 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
292 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
415 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
251 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
362 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
363 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
251 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
360 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
361 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c353 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
355 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
532 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
533 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
652 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
653 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
314 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
439 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
440 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
549 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
550 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
314 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
439 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
440 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
549 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
550 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c351 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
353 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
528 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
529 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
647 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
648 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
314 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
442 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
443 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
554 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
555 vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
355 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
533 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
534 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
653 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
654 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
314 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
442 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
443 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
554 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
555 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
355 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
533 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
534 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
653 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
654 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c353 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
355 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
535 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
536 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
657 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
658 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
355 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
536 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
537 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
658 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
659 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
355 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
536 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
537 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
658 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
659 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()