Home
last modified time | relevance | path

Searched refs:vk18x0123 (Results 1 – 25 of 61) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-wasmrelaxedsimd-fma.c335 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
337 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
521 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
522 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
645 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
646 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
Dup8x25-wasmrelaxedsimd-fma.c333 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
335 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
516 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
517 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
639 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
640 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
Dup8x25-minmax-wasmrelaxedsimd-fma-acc2.c335 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
337 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
524 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
525 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
650 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
651 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
Dup4x25-minmax-neonfma.c242 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
243 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
352 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
353 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon-acc2.c242 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
243 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
354 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
355 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma-acc2.c242 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
243 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
354 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
355 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmrelaxedsimd-fma.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
280 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
404 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
405 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
Dup4x25-wasmrelaxedsimd-fma.c277 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
278 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
400 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
401 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
Dup4x25-minmax-wasmrelaxedsimd.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
404 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
405 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
Dup4x25-minmax-neon.c242 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
243 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
352 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
353 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-minmax-wasmrelaxedsimd-fma-acc2.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
280 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
406 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
407 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
Dup4x25-minmax-sse-acc2.c279 const __m128 vk18x0123 = _mm_load_ps(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
280 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
405 const __m128 vk18x0123 = _mm_load_ps(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
406 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-sse.c279 const __m128 vk18x0123 = _mm_load_ps(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
280 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
403 const __m128 vk18x0123 = _mm_load_ps(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
404 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-wasmsimd-arm-acc2.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
406 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
407 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-wasmsimd-x86-acc2.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
406 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
407 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-wasmsimd-arm.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
404 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
405 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-wasmrelaxedsimd-acc2.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
406 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
407 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
Dup4x25-minmax-wasmsimd-x86.c279 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
404 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
405 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-wasmsimd.c277 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
278 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
400 const v128_t vk18x0123 = wasm_v128_load(w + 76); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
401 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup8x25-minmax-sse-acc2.c335 const __m128 vk18x0123 = _mm_load_ps(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
337 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
523 const __m128 vk18x0123 = _mm_load_ps(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
524 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
649 const __m128 vk18x0123 = _mm_load_ps(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
650 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-neonfma-acc2.c298 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
300 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
434 const float32x4_t vk18x0123 = vld1q_f32(w + 148); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
435 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
546 const float32x4_t vk18x0123 = vld1q_f32(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
547 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-neonfma.c298 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
300 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
431 const float32x4_t vk18x0123 = vld1q_f32(w + 148); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
432 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
541 const float32x4_t vk18x0123 = vld1q_f32(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
542 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-neon.c298 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
300 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
431 const float32x4_t vk18x0123 = vld1q_f32(w + 148); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
432 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
541 const float32x4_t vk18x0123 = vld1q_f32(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
542 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neon-acc2.c298 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
300 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
434 const float32x4_t vk18x0123 = vld1q_f32(w + 148); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
435 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
546 const float32x4_t vk18x0123 = vld1q_f32(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
547 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c335 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
337 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
524 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
525 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
650 const v128_t vk18x0123 = wasm_v128_load(w + 152); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
651 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()

123