Home
last modified time | relevance | path

Searched refs:vk15x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c261 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
262 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
392 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
393 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c261 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
262 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
394 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
395 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c261 const __m128 vk15x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
262 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
391 const __m128 vk15x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
392 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c261 const __m128 vk15x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
262 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
393 const __m128 vk15x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
394 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c259 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
260 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
388 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
389 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c230 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
231 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
340 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
341 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c261 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
262 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
392 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
393 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c230 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
231 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
342 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
343 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c261 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
262 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
394 const v128_t vk15x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
395 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c230 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
231 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
342 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
343 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c230 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
231 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
340 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
341 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c308 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
310 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
502 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
503 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
632 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
633 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c277 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
279 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
419 const float32x4_t vk15x0123 = vld1q_f32(w + 124); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
420 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
529 const float32x4_t vk15x0123 = vld1q_f32(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
530 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c277 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
279 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
419 const float32x4_t vk15x0123 = vld1q_f32(w + 124); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
420 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
529 const float32x4_t vk15x0123 = vld1q_f32(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
530 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c306 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
308 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
498 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
499 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
627 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
628 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c277 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
279 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
422 const float32x4_t vk15x0123 = vld1q_f32(w + 124); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
423 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
534 const float32x4_t vk15x0123 = vld1q_f32(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
535 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c308 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
310 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
503 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
504 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
633 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
634 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c277 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
279 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
422 const float32x4_t vk15x0123 = vld1q_f32(w + 124); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
423 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
534 const float32x4_t vk15x0123 = vld1q_f32(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
535 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c308 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
310 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
503 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
504 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
633 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
634 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c308 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
310 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
505 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
506 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
637 const __m128 vk15x0123 = _mm_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
638 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c308 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
506 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
507 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
638 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
639 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c308 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
506 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
507 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
638 const v128_t vk15x0123 = wasm_v128_load(w + 128); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
639 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()