Re: [dpdk-dev] [PATCH v3 2/2] net: add support for AVX512/VPCLMULQDQ based CRC

["De Lara Guarch, Pablo" <pablo.de.lara.guarch@intel.com>]

Hi Mairtin,

> -----Original Message-----
> From: O'loingsigh, Mairtin <mairtin.oloingsigh@intel.com>
> Sent: Tuesday, September 29, 2020 4:36 PM
> To: Singh, Jasvinder <jasvinder.singh@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>; De Lara Guarch, Pablo
> <pablo.de.lara.guarch@intel.com>
> Cc: dev@dpdk.org; Ryan, Brendan <brendan.ryan@intel.com>; Coyle, David
> <david.coyle@intel.com>; O'loingsigh, Mairtin <mairtin.oloingsigh@intel.com>
> Subject: [PATCH v3 2/2] net: add support for AVX512/VPCLMULQDQ based CRC
> 
> This patch enables the optimized calculation of CRC32-Ethernet and CRC16-
> CCITT using the AVX512 and VPCLMULQDQ instruction sets. This CRC
> implementation is built if the compiler supports the required instruction sets. It is
> selected at run-time if the host CPU, again, supports the required instruction
> sets.
> 
> Signed-off-by: Mairtin o Loingsigh <mairtin.oloingsigh@intel.com>
> Signed-off-by: David Coyle <david.coyle@intel.com>

...

> +static __rte_always_inline uint32_t
> +crc32_eth_calc_vpclmulqdq(const uint8_t *data, uint32_t data_len, uint32_t
> crc,
> +	const struct crc_vpclmulqdq_ctx *params) {
> +	__m128i res, d;
> +	__m256i b;
> +	__m512i temp, k;
> +	__m512i qw0 = _mm512_set1_epi64(0), qw1, qw2, qw3;
> +	__m512i fold0, fold1, fold2, fold3;
> +	__mmask16 mask;
> +	uint32_t n = 0;
> +	int reduction = 0;
> +
> +	/* Get CRC init value */
> +	b = _mm256_insert_epi32(_mm256_setzero_si256(), crc, 0);
> +	temp = _mm512_inserti32x8(_mm512_setzero_si512(), b, 0);

You can replace this with the following, which produces less instructions
(b needs to be changed to __m128i):

        b = _mm_cvtsi32_si128(crc);
        temp = _mm512_castsi128_si512(b);

> +
> +	if (data_len > 255) {
> +		fold0 = _mm512_loadu_si512((const __m512i *)data);

...

> +	} else {
> +		if (data_len > 31) {
> +			res = _mm_insert_epi32(_mm_setzero_si128(), crc, 0);

Should work better with:

res = _mm_cvtsi32_si128(crc);

> +			d = _mm_loadu_si128((const __m128i *)data);
> +			res = _mm_xor_si128(res, d);
> +			n += 16;
> +
> +			reduction = 240 - ((n+256)-data_len);
> +
> +			while (reduction > 0)
> +				reduction_loop(&res, &reduction, data, &n,
> +						params);
> +
> +			if (n != data_len)
> +				res = last_two_xmm(data, data_len, n, res,
> +						params);
> +		} else if (data_len > 16) {
> +			res = _mm_insert_epi32(_mm_setzero_si128(), crc, 0);

Same as above.

> +			d = _mm_loadu_si128((const __m128i *)data);
> +			res = _mm_xor_si128(res, d);
> +			n += 16;
> +
> +			if (n != data_len)
> +				res = last_two_xmm(data, data_len, n, res,
> +						params);
> +		} else if (data_len == 16) {
> +			res = _mm_insert_epi32(_mm_setzero_si128(), crc, 0);

Same.

> +			d = _mm_loadu_si128((const __m128i *)data);
> +			res = _mm_xor_si128(res, d);
> +		} else {
> +			res = _mm_insert_epi32(_mm_setzero_si128(), crc, 0);

Same.

> +			mask = byte_len_to_mask_table[data_len];
> +			d = _mm_maskz_loadu_epi8(mask, data);