Home
last modified time | relevance | path

Searched refs:vi15x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c258 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
391 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c258 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
393 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c258 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
390 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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.c258 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
392 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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.c256 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
387 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c229 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
339 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c258 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
391 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c229 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
341 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c258 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
393 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c229 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
341 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c229 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
339 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c304 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
499 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
631 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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.c275 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
418 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
420 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
528 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c275 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
418 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
420 vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
528 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c302 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
495 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
626 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c275 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
421 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; 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()
533 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c304 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
500 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
632 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c275 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 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()
421 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; 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()
533 const float32x4_t vi15x0123 = vld1q_f32(i15); 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.c304 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
500 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
632 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c304 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
502 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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()
636 const __m128 vi15x0123 = _mm_loadu_ps(i15); 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.c304 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
503 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
637 const v128_t vi15x0123 = wasm_v128_load(i15); 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.c304 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
503 const v128_t vi15x0123 = wasm_v128_load(i15); 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()
637 const v128_t vi15x0123 = wasm_v128_load(i15); 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()