Home
last modified time | relevance | path

Searched refs:vk24x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c315 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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()
430 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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.c315 const __m128 vk24x0123 = _mm_load_ps(w + 100); 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()
427 const __m128 vk24x0123 = _mm_load_ps(w + 100); 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.c266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
378 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; 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.c266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
376 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; 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.c266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
376 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; 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.c313 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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()
424 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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.c266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
378 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; 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.c315 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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()
428 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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.c315 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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()
430 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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.c315 const __m128 vk24x0123 = _mm_load_ps(w + 100); 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()
429 const __m128 vk24x0123 = _mm_load_ps(w + 100); 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.c315 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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()
428 const v128_t vk24x0123 = wasm_v128_load(w + 100); 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.c340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
458 const float32x4_t vk24x0123 = vld1q_f32(w + 196); 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()
570 const float32x4_t vk24x0123 = vld1q_f32(w + 200); 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.c389 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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()
556 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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()
668 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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.c389 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
557 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
669 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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.c340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
455 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
456 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
565 const float32x4_t vk24x0123 = vld1q_f32(w + 200); 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.c389 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
560 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
674 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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.c389 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
560 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
674 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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.c340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
455 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
456 vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
565 const float32x4_t vk24x0123 = vld1q_f32(w + 200); 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.c389 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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()
559 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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()
673 const __m128 vk24x0123 = _mm_load_ps(w + 200); 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.c387 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
552 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
663 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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.c389 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
557 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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()
669 const v128_t vk24x0123 = wasm_v128_load(w + 200); 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.c340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 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()
458 const float32x4_t vk24x0123 = vld1q_f32(w + 196); 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()
570 const float32x4_t vk24x0123 = vld1q_f32(w + 200); 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()