// Auto-generated file. Do not edit! // Template: src/qs8-f32-vcvt/avx2.c.in // Generator: tools/xngen // // Copyright 2021 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include "xnnpack/common.h" #include "xnnpack/intrinsics-polyfill.h" #include "xnnpack/vcvt.h" void xnn_qu8_f32_vcvt_ukernel__avx2_u32( size_t batch, const uint8_t* input, float* output, const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(batch != 0); assert(batch % sizeof(uint8_t) == 0); assert(input != NULL); assert(output != NULL); const __m256i vzero_point = _mm256_set1_epi32(params->scalar.zero_point); const __m256 vscale = _mm256_set1_ps(params->scalar.scale); XNN_FORCE_REALIZATION(vzero_point); XNN_FORCE_REALIZATION(vscale); for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) { __m256i vx01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input)); __m256i vx89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input + 8))); __m256i vxGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input + 16))); __m256i vxOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input + 24))); input += 32; vx01234567 = _mm256_sub_epi32(vx01234567, vzero_point); vx89ABCDEF = _mm256_sub_epi32(vx89ABCDEF, vzero_point); vxGHIJKLMN = _mm256_sub_epi32(vxGHIJKLMN, vzero_point); vxOPQRSTUV = _mm256_sub_epi32(vxOPQRSTUV, vzero_point); __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567); __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); __m256 vyGHIJKLMN = _mm256_cvtepi32_ps(vxGHIJKLMN); __m256 vyOPQRSTUV = _mm256_cvtepi32_ps(vxOPQRSTUV); vy01234567 = _mm256_mul_ps(vy01234567, vscale); vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); vyGHIJKLMN = _mm256_mul_ps(vyGHIJKLMN, vscale); vyOPQRSTUV = _mm256_mul_ps(vyOPQRSTUV, vscale); _mm256_storeu_ps(output, vy01234567); _mm256_storeu_ps(output + 8, vy89ABCDEF); _mm256_storeu_ps(output + 16, vyGHIJKLMN); _mm256_storeu_ps(output + 24, vyOPQRSTUV); output += 32; } for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { __m256i vx = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input)); vx = _mm256_sub_epi32(vx, vzero_point); input += 8; __m256 vy = _mm256_cvtepi32_ps(vx); vy = _mm256_mul_ps(vy, vscale); _mm256_storeu_ps(output, vy); output += 8; } if XNN_UNLIKELY(batch != 0) { assert(batch >= 1 * sizeof(uint8_t)); assert(batch <= 7 * sizeof(uint8_t)); __m256i vx = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input)); vx = _mm256_sub_epi32(vx, vzero_point); __m256 vy = _mm256_cvtepi32_ps(vx); vy = _mm256_mul_ps(vy, vscale); __m128 vy_lo = _mm256_castps256_ps128(vy); if (batch & (4 * sizeof(uint8_t))) { _mm_storeu_ps(output, vy_lo); vy_lo = _mm256_extractf128_ps(vy, 1); output += 4; } if (batch & (2 * sizeof(uint8_t))) { _mm_storel_pi((__m64*) output, vy_lo); vy_lo = _mm_movehl_ps(vy_lo, vy_lo); output += 2; } if (batch & (1 * sizeof(uint8_t))) { _mm_store_ss(output, vy_lo); } } }