<div dir="ltr"><div dir="ltr"> </div> <div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Tue, Nov 10, 2026 at 3:21 PM Konstantin Ananyev <<a href="mailto:konstantin.v.ananyev@yandex.ru">konstantin.v.ananyev@yandex.ru</a>> wrote: </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> > > > On Mon, Nov 9, 2026 at 3:40 PM Konstantin Ananyev > <<a href="mailto:konstantin.v.ananyev@yandex.ru" target="_blank">konstantin.v.ananyev@yandex.ru</a>> wrote: > > > > Replace the __m128i members in the PCLMULQDQ CRC context with > aligned > > uint64_t arrays to allow portable compile-time initialization. > > > > Move CRC16-CCITT and CRC32-ETH folding constants into static > const context > > structures in their respective handlers, removing the runtime > > initialization previously performed by rte_net_crc_sse42_init > and avoids > > the extra data segment static variables which were being populated. > > > > Load the constants into SIMD registers using aligned loads when > performing > > CRC calculations. This was done transparently previously when the > > __m128i values were assigned, now there is explicit SIMD load done. > > It is not clear why(how) new code is better then what we have > right now. > Can you probably provide a bit more details? > > > Previously this file was maintaining static variables in data segment > namely: > static alignas(16) struct crc_pclmulqdq_ctx crc32_eth_pclmulqdq; > static alignas(16) struct crc_pclmulqdq_ctx crc16_ccitt_pclmulqdq; > > These were populated in the init function with the same calculated > fold constants. > The constants already are present in the rodata section in the binary. > Hence the init is doing extra work in copying it over to the > static crc_pclmulqdq_ctx > structures which is redundant. > > This patch avoids this extra copy during init and modifies the SIMD > routines > to directly read the constants from rodata memory. Thus the data > segment memory usage is > reduced by getting rid of the static variables (128 bytes). There > should be no > performance impact by this change (minus the init work done previously). TBH, it looks to me like a nitpicking: yes we move 128B from .data to .rodata, but in terms of functionality or perfomance - does it really matter? </blockquote><div> </div><div>It is not moving from .data to .rodata, same data was already there in .rodata,</div><div>hence it is removal from .data.</div><div> </div><div>Well this was mainly motivated from the rewrite for improving barrett reduction.</div><div>I've combined this in the other patch and we can drop this patch from review.</div><div>Kindly review the newer submission which includes this refactory change as well:</div><div><a href="https://patches.dpdk.org/project/dpdk/patch/20261110121048.831330-1-16567adigashreesh@gmail.com/">https://patches.dpdk.org/project/dpdk/patch/20261110121048.831330-1-16567adigashreesh@gmail.com/</a></div><div> </div><div>Thanks,</div><div>Shreesh</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> Though yes, the code looks a bit simpler and nicer. If we decided to improve the init code, then my preference would be to use rte_xmm_t? Then we can do something like that: struct crc_pclmulqdq_ctx { rte_xmm_t rk1_rk2; ... }; ... static const struct crc_pclmulqdq_ctx crc32_eth_pclmulqdq = { .rk1_rk2 = { .u64 = {0x1c6e41596LLU, 0x154442bd4LLU }}, ... }; ... __m128i k = params->rk3_rk4.x; > > > > > Signed-off-by: Shreesh Adiga <<a href="mailto:16567adigashreesh@gmail.com" target="_blank">16567adigashreesh@gmail.com</a>> > > --- > > lib/net/net_crc_sse.c | 68 > ++++++++++++++----------------------------- > > 1 file changed, 22 insertions(+), 46 deletions(-) > > > > diff --git a/lib/net/net_crc_sse.c b/lib/net/net_crc_sse.c > > index e30f8544fc..4361e35e39 100644 > > --- a/lib/net/net_crc_sse.c > > +++ b/lib/net/net_crc_sse.c > > @@ -12,15 +12,13 @@ > > #include "net_crc.h" > > > > /** PCLMULQDQ CRC computation context structure */ > > -struct crc_pclmulqdq_ctx { > > - __m128i rk1_rk2; > > - __m128i rk3_rk4; > > - __m128i rk5_rk6; > > - __m128i rk7_rk8; > > +struct __rte_aligned(XMM_SIZE) crc_pclmulqdq_ctx { > > + uint64_t rk1_rk2[2]; > > + uint64_t rk3_rk4[2]; > > + uint64_t rk5_rk6[2]; > > + uint64_t rk7_rk8[2]; > > }; > > > > -static alignas(16) struct crc_pclmulqdq_ctx crc32_eth_pclmulqdq; > > -static alignas(16) struct crc_pclmulqdq_ctx crc16_ccitt_pclmulqdq; > > /** > > * @brief Performs one folding round > > * > > @@ -162,7 +160,7 @@ crc32_eth_calc_pclmulqdq( > > fold3 = _mm_loadu_si128((const __m128i *)(data + 32)); > > fold4 = _mm_loadu_si128((const __m128i *)(data + 48)); > > fold1 = _mm_xor_si128(fold1, temp); > > - k = params->rk1_rk2; > > + k = _mm_load_si128((const __m128i *)params->rk1_rk2); > > > > for (n = 64; (n + 64) <= data_len; n += 64) { > > temp1 = _mm_loadu_si128((const __m128i > *)&data[n]); > > @@ -175,7 +173,7 @@ crc32_eth_calc_pclmulqdq( > > fold4 = crcr32_folding_round(temp4, k, fold4); > > } > > > > - k = params->rk3_rk4; > > + k = _mm_load_si128((const __m128i *)params->rk3_rk4); > > fold1 = crcr32_folding_round(fold2, k, fold1); > > fold1 = crcr32_folding_round(fold3, k, fold1); > > fold = crcr32_folding_round(fold4, k, fold1); > > @@ -205,7 +203,7 @@ crc32_eth_calc_pclmulqdq( > > fold = _mm_xor_si128(fold, temp); > > > > /** Single folding loop - the last 16 bytes is processed > separately */ > > - k = params->rk3_rk4; > > + k = _mm_load_si128((const __m128i *)params->rk3_rk4); > > n = 16; > > > > single_fold_loop: > > @@ -235,11 +233,11 @@ crc32_eth_calc_pclmulqdq( > > > > /** Reduction 128 -> 32 Assumes: fold holds 128bit folded > data */ > > reduction_128_64: > > - k = params->rk5_rk6; > > + k = _mm_load_si128((const __m128i *)params->rk5_rk6); > > fold = crcr32_reduce_128_to_64(fold, k); > > > > barret_reduction: > > - k = params->rk7_rk8; > > + k = _mm_load_si128((const __m128i *)params->rk7_rk8); > > n = crcr32_reduce_64_to_32(fold, k); > > > > return n; > > @@ -248,45 +246,17 @@ crc32_eth_calc_pclmulqdq( > > void > > rte_net_crc_sse42_init(void) > > { > > - uint64_t k1, k2, k3, k4, k5, k6; > > - uint64_t p = 0, q = 0; > > - > > - /** Initialize CRC16 data */ > > - k1 = 0x14ff2LLU; > > - k2 = 0x19a3cLLU; > > - k3 = 0x189aeLLU; > > - k4 = 0x8e10LLU; > > - k5 = 0x189aeLLU; > > - k6 = 0x114aaLLU; > > - q = 0x11c581910LLU; > > - p = 0x10811LLU; > > - > > - /** Save the params in context structure */ > > - crc16_ccitt_pclmulqdq.rk1_rk2 = _mm_set_epi64x(k2, k1); > > - crc16_ccitt_pclmulqdq.rk3_rk4 = _mm_set_epi64x(k4, k3); > > - crc16_ccitt_pclmulqdq.rk5_rk6 = _mm_set_epi64x(k6, k5); > > - crc16_ccitt_pclmulqdq.rk7_rk8 = _mm_set_epi64x(p, q); > > - > > - /** Initialize CRC32 data */ > > - k1 = 0x1c6e41596LLU; > > - k2 = 0x154442bd4LLU; > > - k3 = 0xccaa009eLLU; > > - k4 = 0x1751997d0LLU; > > - k5 = 0xccaa009eLLU; > > - k6 = 0x163cd6124LLU; > > - q = 0x1f7011640LLU; > > - p = 0x1db710641LLU; > > - > > - /** Save the params in context structure */ > > - crc32_eth_pclmulqdq.rk1_rk2 = _mm_set_epi64x(k2, k1); > > - crc32_eth_pclmulqdq.rk3_rk4 = _mm_set_epi64x(k4, k3); > > - crc32_eth_pclmulqdq.rk5_rk6 = _mm_set_epi64x(k6, k5); > > - crc32_eth_pclmulqdq.rk7_rk8 = _mm_set_epi64x(p, q); > > } > > > > uint32_t > > rte_crc16_ccitt_sse42_handler(const uint8_t *data, uint32_t > data_len) > > { > > + static const struct crc_pclmulqdq_ctx > crc16_ccitt_pclmulqdq = { > > + .rk1_rk2 = { 0x14ff2LLU, 0x19a3cLLU }, > > + .rk3_rk4 = { 0x189aeLLU, 0x8e10LLU }, > > + .rk5_rk6 = { 0x189aeLLU, 0x114aaLLU }, > > + .rk7_rk8 = { 0x11c581910LLU, 0x10811LLU }, > > + }; > > /** return 16-bit CRC value */ > > return (uint16_t)~crc32_eth_calc_pclmulqdq(data, > > data_len, > > @@ -297,6 +267,12 @@ rte_crc16_ccitt_sse42_handler(const uint8_t > *data, uint32_t data_len) > > uint32_t > > rte_crc32_eth_sse42_handler(const uint8_t *data, uint32_t > data_len) > > { > > + static const struct crc_pclmulqdq_ctx crc32_eth_pclmulqdq = { > > + .rk1_rk2 = { 0x1c6e41596LLU, 0x154442bd4LLU }, > > + .rk3_rk4 = { 0xccaa009eLLU, 0x1751997d0LLU }, > > + .rk5_rk6 = { 0xccaa009eLLU, 0x163cd6124LLU }, > > + .rk7_rk8 = { 0x1f7011640LLU, 0x1db710641LLU }, > > + }; > > return ~crc32_eth_calc_pclmulqdq(data, > > data_len, > > 0xffffffffUL, > </blockquote></div></div>