Home
last modified time | relevance | path

Searched refs:vk9x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
226 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
370 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
371 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c225 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
226 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
367 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
368 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
207 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
318 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
319 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
207 vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
316 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
317 vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
207 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
316 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
317 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c223 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
224 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
364 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
365 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
207 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
318 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
319 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
226 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
368 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
369 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
226 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
370 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
371 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c225 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
226 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
369 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
370 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
226 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
368 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
369 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
237 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
398 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
399 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
510 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
511 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c254 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
256 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
466 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
467 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
608 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
609 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
467 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
468 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
609 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
610 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
237 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
395 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
396 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
505 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
506 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
256 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
470 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
471 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
614 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
615 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
256 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
470 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
471 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
614 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
615 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
237 vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
395 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
396 vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
505 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
506 vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c254 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
256 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
469 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
470 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
613 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
614 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c252 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
254 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
462 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
463 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
603 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
604 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
467 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
468 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
609 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
610 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
237 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
398 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
399 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
510 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
511 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()