Home
last modified time | relevance | path

Searched refs:vk13x0123 (Results 1 – 25 of 61) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-wasmrelaxedsimd-fma.c290 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
292 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
491 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
492 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
625 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
626 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma()
Dup8x25-wasmrelaxedsimd-fma.c288 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
290 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
486 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
487 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
619 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
620 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma()
Dup8x25-minmax-wasmrelaxedsimd-fma-acc2.c290 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
292 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
494 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
495 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
630 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2() local
631 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma_acc2()
Dup4x25-minmax-neonfma.c222 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
223 vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
332 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
333 vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon-acc2.c222 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
223 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
334 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
335 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma-acc2.c222 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
223 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
334 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
335 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmrelaxedsimd-fma.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
250 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
384 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
385 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma()
Dup4x25-wasmrelaxedsimd-fma.c247 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
248 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
380 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
381 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma()
Dup4x25-minmax-wasmrelaxedsimd.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
250 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
384 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
385 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd()
Dup4x25-minmax-neon.c222 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
223 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
332 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
333 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-minmax-wasmrelaxedsimd-fma-acc2.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
250 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
386 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
387 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2()
Dup4x25-minmax-sse-acc2.c249 const __m128 vk13x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
250 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
385 const __m128 vk13x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
386 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-sse.c249 const __m128 vk13x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
250 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
383 const __m128 vk13x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
384 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-wasmsimd-arm-acc2.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
386 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
387 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-wasmsimd-x86-acc2.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
386 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
387 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-wasmsimd-arm.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
250 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
384 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
385 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-wasmrelaxedsimd-acc2.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
386 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
387 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2()
Dup4x25-minmax-wasmsimd-x86.c249 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
250 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
384 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
385 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-wasmsimd.c247 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
248 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
380 const v128_t vk13x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
381 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup8x25-minmax-sse-acc2.c290 const __m128 vk13x0123 = _mm_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
292 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
493 const __m128 vk13x0123 = _mm_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
494 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
629 const __m128 vk13x0123 = _mm_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
630 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-neonfma-acc2.c263 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
265 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
414 const float32x4_t vk13x0123 = vld1q_f32(w + 108); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
415 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
526 const float32x4_t vk13x0123 = vld1q_f32(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
527 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-neonfma.c263 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
265 vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
411 const float32x4_t vk13x0123 = vld1q_f32(w + 108); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
412 vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
521 const float32x4_t vk13x0123 = vld1q_f32(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
522 vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-neon.c263 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
265 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
411 const float32x4_t vk13x0123 = vld1q_f32(w + 108); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
412 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
521 const float32x4_t vk13x0123 = vld1q_f32(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
522 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neon-acc2.c263 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
265 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
414 const float32x4_t vk13x0123 = vld1q_f32(w + 108); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
415 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
526 const float32x4_t vk13x0123 = vld1q_f32(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
527 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c290 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
292 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
494 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
495 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
630 const v128_t vk13x0123 = wasm_v128_load(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
631 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()

123