/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-minmax-wasmsimd-x86.c | 276 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 403 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 276 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 405 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|
D | up4x25-minmax-sse.c | 276 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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() 402 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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()
|
D | up4x25-minmax-sse-acc2.c | 276 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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() 404 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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()
|
D | up4x25-wasmsimd.c | 274 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 399 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|
D | up4x25-minmax-neonfma.c | 241 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 351 const float32x4_t vi18x0123 = vld1q_f32(i18); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 353 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
|
D | up4x25-minmax-wasmsimd-arm.c | 276 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 403 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|
D | up4x25-minmax-neonfma-acc2.c | 241 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 353 const float32x4_t vi18x0123 = vld1q_f32(i18); 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()
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 276 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 405 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|
D | up4x25-minmax-neon-acc2.c | 241 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 353 const float32x4_t vi18x0123 = vld1q_f32(i18); 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()
|
D | up4x25-minmax-neon.c | 241 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 351 const float32x4_t vi18x0123 = vld1q_f32(i18); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 353 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
|
D | up8x25-minmax-sse.c | 331 const __m128 vi18x0123 = _mm_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 337 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() 517 const __m128 vi18x0123 = _mm_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 521 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() 643 const __m128 vi18x0123 = _mm_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 645 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
|
D | up8x25-minmax-neon.c | 296 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 430 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 432 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() 540 const float32x4_t vi18x0123 = vld1q_f32(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 542 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
|
D | up8x25-minmax-neonfma.c | 296 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 430 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 432 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() 540 const float32x4_t vi18x0123 = vld1q_f32(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 542 vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
|
D | up8x25-wasmsimd.c | 329 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 335 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() 513 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 517 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() 638 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 640 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
|
D | up8x25-minmax-neonfma-acc2.c | 296 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 433 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; 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() 545 const float32x4_t vi18x0123 = vld1q_f32(i18); 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()
|
D | up8x25-minmax-wasmsimd-arm.c | 331 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 337 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() 518 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 522 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() 644 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 646 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
|
D | up8x25-minmax-neon-acc2.c | 296 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 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() 433 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; 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() 545 const float32x4_t vi18x0123 = vld1q_f32(i18); 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()
|
D | up8x25-minmax-wasmsimd-x86.c | 331 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 337 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() 518 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 522 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() 644 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 646 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
|
D | up8x25-minmax-sse-acc2.c | 331 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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() 520 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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() 648 const __m128 vi18x0123 = _mm_loadu_ps(i18); 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()
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 331 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 337 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() 521 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 525 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() 649 const v128_t vi18x0123 = wasm_v128_load(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 651 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 331 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 521 const v128_t vi18x0123 = wasm_v128_load(i18); 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() 649 const v128_t vi18x0123 = wasm_v128_load(i18); 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()
|