// Copyright 2023 The Wuffs Authors. // // Licensed under the Apache License, Version 2.0 or the MIT license // , at your // option. This file may not be copied, modified, or distributed // except according to those terms. // // SPDX-License-Identifier: Apache-2.0 OR MIT pri func decoder.decode_idct_x86_avx2!(dst_buffer: slice base.u8, dst_stride: base.u64, q: base.u32[..= 3]), choose cpu_arch >= x86_avx2, { var sse42_util : base.x86_sse42_utility var util : base.x86_avx2_utility // SIMD constants. var k_0000 : base.x86_m256i var k_8080 : base.x86_m256i var k_0000_0002 : base.x86_m256i var k_0001_FFFF : base.x86_m256i var k_0400_0000 : base.x86_m256i var k_29CF_1151_D630_1151 : base.x86_m256i var k_E333_133E_ADFD_1051 : base.x86_m256i var k_E6DC_25A1_1925_25A1 : base.x86_m256i var k_ECC1_E333_EFB0_ADFD : base.x86_m256i // "Are the AC terms zero" variable names start with "az_". var az_coeffs : base.x86_m128i var az_ah00 : base.x86_m256i var az_ad00 : base.x86_m256i var az_eh00 : base.x86_m256i var az_adeh : base.x86_m256i // Set up variables. var rows01 : base.x86_m256i var rows23 : base.x86_m256i var rows45 : base.x86_m256i var rows67 : base.x86_m256i var quants01 : base.x86_m256i var quants23 : base.x86_m256i var quants45 : base.x86_m256i var quants67 : base.x86_m256i var rows04 : base.x86_m256i var rows31 : base.x86_m256i var rows26 : base.x86_m256i var rows75 : base.x86_m256i // First pass variable names start with "fp_". var fp_rows62 : base.x86_m256i var fp_bq2662ad : base.x86_m256i var fp_bq2662eh : base.x86_m256i var fp_cb26ad : base.x86_m256i var fp_cb26eh : base.x86_m256i var fp_rows40pos : base.x86_m256i var fp_rows04neg : base.x86_m256i var fp_rows0pm4 : base.x86_m256i var fp_ccpmad : base.x86_m256i var fp_ccpmeh : base.x86_m256i var fp_cd01ad : base.x86_m256i var fp_cd01eh : base.x86_m256i var fp_cd32ad : base.x86_m256i var fp_cd32eh : base.x86_m256i var fp_sums7351 : base.x86_m256i var fp_sums5173 : base.x86_m256i var fp_ci73515173ad : base.x86_m256i var fp_ci73515173eh : base.x86_m256i var fp_cl7351ad : base.x86_m256i var fp_cl7351eh : base.x86_m256i var fp_rows13 : base.x86_m256i var fp_bq7153ad : base.x86_m256i var fp_bq7153eh : base.x86_m256i var fp_ck75ad : base.x86_m256i var fp_ck75eh : base.x86_m256i var fp_cl5173ad : base.x86_m256i var fp_cl5173eh : base.x86_m256i var fp_ck13ad : base.x86_m256i var fp_ck13eh : base.x86_m256i // Intermediate variables between first and second pass. var intermediate01ad : base.x86_m256i var intermediate01eh : base.x86_m256i var intermediate01 : base.x86_m256i var intermediate32ad : base.x86_m256i var intermediate32eh : base.x86_m256i var intermediate32 : base.x86_m256i var intermediate45ad : base.x86_m256i var intermediate45eh : base.x86_m256i var intermediate45 : base.x86_m256i var intermediate76ad : base.x86_m256i var intermediate76eh : base.x86_m256i var intermediate76 : base.x86_m256i var ita0a1e0e1 : base.x86_m256i var ita2a3e2e3 : base.x86_m256i var ita4a5e4e5 : base.x86_m256i var ita6a7e6e7 : base.x86_m256i var ita0c0e0g0 : base.x86_m256i var ita1c1e1g1 : base.x86_m256i var ita4c4e4g4 : base.x86_m256i var ita5c5e5g5 : base.x86_m256i var ita0b0e0f0 : base.x86_m256i var ita4b4e4f4 : base.x86_m256i var itc0d0g0h0 : base.x86_m256i var itc4d4g4h4 : base.x86_m256i var intermediateae : base.x86_m256i var intermediatebf : base.x86_m256i var intermediatecg : base.x86_m256i var intermediatedh : base.x86_m256i var intermediatedb : base.x86_m256i var intermediatehf : base.x86_m256i // Second pass variable names start with "sp_". var sp_cols62 : base.x86_m256i var sp_bq2662ad : base.x86_m256i var sp_bq2662eh : base.x86_m256i var sp_rb26ad : base.x86_m256i var sp_rb26eh : base.x86_m256i var sp_cols40pos : base.x86_m256i var sp_cols04neg : base.x86_m256i var sp_cols0pm4 : base.x86_m256i var sp_rcpmad : base.x86_m256i var sp_rcpmeh : base.x86_m256i var sp_rd01ad : base.x86_m256i var sp_rd01eh : base.x86_m256i var sp_rd32ad : base.x86_m256i var sp_rd32eh : base.x86_m256i var sp_sums7351 : base.x86_m256i var sp_sums5173 : base.x86_m256i var sp_ri73515173ad : base.x86_m256i var sp_ri73515173eh : base.x86_m256i var sp_rl7351ad : base.x86_m256i var sp_rl7351eh : base.x86_m256i var sp_cols13 : base.x86_m256i var sp_bq7153ad : base.x86_m256i var sp_bq7153eh : base.x86_m256i var sp_rk75ad : base.x86_m256i var sp_rk75eh : base.x86_m256i var sp_rl5173ad : base.x86_m256i var sp_rl5173eh : base.x86_m256i var sp_rk13ad : base.x86_m256i var sp_rk13eh : base.x86_m256i // Final variables. var final01ad : base.x86_m256i var final01eh : base.x86_m256i var final01 : base.x86_m256i var final32ad : base.x86_m256i var final32eh : base.x86_m256i var final32 : base.x86_m256i var final45ad : base.x86_m256i var final45eh : base.x86_m256i var final45 : base.x86_m256i var final76ad : base.x86_m256i var final76eh : base.x86_m256i var final76 : base.x86_m256i var fta0a1e0e1 : base.x86_m256i var fta2a3e2e3 : base.x86_m256i var fta4a5e4e5 : base.x86_m256i var fta6a7e6e7 : base.x86_m256i var fta0c0e0g0 : base.x86_m256i var fta1c1e1g1 : base.x86_m256i var fta4c4e4g4 : base.x86_m256i var fta5c5e5g5 : base.x86_m256i var fta0b0e0f0 : base.x86_m256i var ftc0d0g0h0 : base.x86_m256i var fta4b4e4f4 : base.x86_m256i var ftc4d4g4h4 : base.x86_m256i var finalae : base.x86_m256i var finalbf : base.x86_m256i var finalcg : base.x86_m256i var finaldh : base.x86_m256i var final0145 : base.x86_m256i var final2367 : base.x86_m256i var final0 : base.u64 var final1 : base.u64 var final2 : base.u64 var final3 : base.u64 var final4 : base.u64 var final5 : base.u64 var final6 : base.u64 var final7 : base.u64 var remaining : slice base.u8 // ---- if 8 > args.dst_stride { return nothing } k_0000 = util.make_m256i_multiple_u16( a00: 0x0000, a01: 0x0000, a02: 0x0000, a03: 0x0000, a04: 0x0000, a05: 0x0000, a06: 0x0000, a07: 0x0000, a08: 0x0000, a09: 0x0000, a10: 0x0000, a11: 0x0000, a12: 0x0000, a13: 0x0000, a14: 0x0000, a15: 0x0000) k_8080 = util.make_m256i_multiple_u16( a00: 0x8080, a01: 0x8080, a02: 0x8080, a03: 0x8080, a04: 0x8080, a05: 0x8080, a06: 0x8080, a07: 0x8080, a08: 0x8080, a09: 0x8080, a10: 0x8080, a11: 0x8080, a12: 0x8080, a13: 0x8080, a14: 0x8080, a15: 0x8080) k_0000_0002 = util.make_m256i_multiple_u16( a00: 0x0000, a01: 0x0002, a02: 0x0000, a03: 0x0002, a04: 0x0000, a05: 0x0002, a06: 0x0000, a07: 0x0002, a08: 0x0000, a09: 0x0002, a10: 0x0000, a11: 0x0002, a12: 0x0000, a13: 0x0002, a14: 0x0000, a15: 0x0002) k_0001_FFFF = util.make_m256i_multiple_u16( a00: 0x0001, a01: 0x0001, a02: 0x0001, a03: 0x0001, a04: 0x0001, a05: 0x0001, a06: 0x0001, a07: 0x0001, a08: 0xFFFF, a09: 0xFFFF, a10: 0xFFFF, a11: 0xFFFF, a12: 0xFFFF, a13: 0xFFFF, a14: 0xFFFF, a15: 0xFFFF) k_0400_0000 = util.make_m256i_multiple_u16( a00: 0x0400, a01: 0x0000, a02: 0x0400, a03: 0x0000, a04: 0x0400, a05: 0x0000, a06: 0x0400, a07: 0x0000, a08: 0x0400, a09: 0x0000, a10: 0x0400, a11: 0x0000, a12: 0x0400, a13: 0x0000, a14: 0x0400, a15: 0x0000) k_29CF_1151_D630_1151 = util.make_m256i_multiple_u16( a00: 0x29CF, a01: 0x1151, a02: 0x29CF, a03: 0x1151, a04: 0x29CF, a05: 0x1151, a06: 0x29CF, a07: 0x1151, a08: 0xD630, a09: 0x1151, a10: 0xD630, a11: 0x1151, a12: 0xD630, a13: 0x1151, a14: 0xD630, a15: 0x1151) k_E333_133E_ADFD_1051 = util.make_m256i_multiple_u16( a00: 0xE333, a01: 0x133E, a02: 0xE333, a03: 0x133E, a04: 0xE333, a05: 0x133E, a06: 0xE333, a07: 0x133E, a08: 0xADFD, a09: 0x1051, a10: 0xADFD, a11: 0x1051, a12: 0xADFD, a13: 0x1051, a14: 0xADFD, a15: 0x1051) k_E6DC_25A1_1925_25A1 = util.make_m256i_multiple_u16( a00: 0xE6DC, a01: 0x25A1, a02: 0xE6DC, a03: 0x25A1, a04: 0xE6DC, a05: 0x25A1, a06: 0xE6DC, a07: 0x25A1, a08: 0x1925, a09: 0x25A1, a10: 0x1925, a11: 0x25A1, a12: 0x1925, a13: 0x25A1, a14: 0x1925, a15: 0x25A1) k_ECC1_E333_EFB0_ADFD = util.make_m256i_multiple_u16( a00: 0xECC1, a01: 0xE333, a02: 0xECC1, a03: 0xE333, a04: 0xECC1, a05: 0xE333, a06: 0xECC1, a07: 0xE333, a08: 0xEFB0, a09: 0xADFD, a10: 0xEFB0, a11: 0xADFD, a12: 0xEFB0, a13: 0xADFD, a14: 0xEFB0, a15: 0xADFD) // ---- while.goto_do_second_pass true, inv 8 <= args.dst_stride, {{ // Name the columns a ..= h and the rows 0 ..= 7. // // Check if the AC terms are zero. First, a relative cheap check of the // (a1, b1, c1, d1, a2, b2, c2, d2) coefficients - two half-rows. if 0 == (util.make_u64_slice_u16lex4(a: this.mcu_blocks[0][0x08 .. 0x0C]) | util.make_u64_slice_u16lex4(a: this.mcu_blocks[0][0x10 .. 0x14])) { // A more thorough check of every coefficient in rows 1 ..= 7. az_coeffs = sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x08 .. 0x10])._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x10 .. 0x18]))._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x18 .. 0x20]))._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x20 .. 0x28]))._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x28 .. 0x30]))._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x30 .. 0x38]))._mm_or_si128(b: sse42_util.make_m128i_slice_u16lex8(a: this.mcu_blocks[0][0x38 .. 0x40])) // Pack u16x8 into u8x8 in the low 64 bits (using signed saturation) // and bulk-compare to zero. Only "zero versus non-zero" matters. if 0 == az_coeffs._mm_packs_epi16(b: az_coeffs).truncate_u64() { // Dequantize, similar to the "Set up" code path below. rows01 = util.make_m256i_slice_u16lex16(a: this.mcu_blocks[0][0x00 .. 0x10]) quants01 = util.make_m256i_slice_u16lex16(a: this.quant_tables[args.q][0x00 .. 0x10]) rows01 = rows01._mm256_mullo_epi16(b: quants01) // az_ah00 = i16x16 [a b c d e f g h 0 0 0 0 0 0 0 0] // az_ad00 = i16x16 [a a b b c c d d 0 0 0 0 0 0 0 0] // az_eh00 = i16x16 [e e f f g g h h 0 0 0 0 0 0 0 0] // az_adeh = i16x16 [a a b b c c d d e e f f g g h h] az_ah00 = rows01._mm256_slli_epi16(imm8: 2) // PASS1_BITS is 2. az_ad00 = az_ah00._mm256_unpacklo_epi16(b: az_ah00) az_eh00 = az_ah00._mm256_unpackhi_epi16(b: az_ah00) az_adeh = az_ad00._mm256_inserti128_si256(b: az_eh00._mm256_castsi256_si128(), imm8: 1) // intermediateae = i16x16 [a a a a a a a a e e e e e e e e] // intermediatebf = i16x16 [b b b b b b b b f f f f f f f f] // intermediatecg = i16x16 [c c c c c c c c g g g g g g g g] // intermediatedh = i16x16 [d d d d d d d d h h h h h h h h] intermediateae = az_adeh._mm256_shuffle_epi32(imm8: 0x00) intermediatebf = az_adeh._mm256_shuffle_epi32(imm8: 0x55) intermediatecg = az_adeh._mm256_shuffle_epi32(imm8: 0xAA) intermediatedh = az_adeh._mm256_shuffle_epi32(imm8: 0xFF) break.goto_do_second_pass } } // ---- // Set up. // Dequantize: // // rows01 = i16x16 [a0 b0 .. g0 h0 // a1 b1 .. g1 h1] // rows23 = i16x16 [a2 b2 .. g2 e2 // a3 b3 .. g3 h3] // rows45 = i16x16 [a4 b4 .. g4 h4 // a5 b5 .. g5 h5] // rows67 = i16x16 [a6 b6 .. g6 h6 // a7 b7 .. g7 h7] rows01 = util.make_m256i_slice_u16lex16(a: this.mcu_blocks[0][0x00 .. 0x10]) rows23 = util.make_m256i_slice_u16lex16(a: this.mcu_blocks[0][0x10 .. 0x20]) rows45 = util.make_m256i_slice_u16lex16(a: this.mcu_blocks[0][0x20 .. 0x30]) rows67 = util.make_m256i_slice_u16lex16(a: this.mcu_blocks[0][0x30 .. 0x40]) quants01 = util.make_m256i_slice_u16lex16(a: this.quant_tables[args.q][0x00 .. 0x10]) quants23 = util.make_m256i_slice_u16lex16(a: this.quant_tables[args.q][0x10 .. 0x20]) quants45 = util.make_m256i_slice_u16lex16(a: this.quant_tables[args.q][0x20 .. 0x30]) quants67 = util.make_m256i_slice_u16lex16(a: this.quant_tables[args.q][0x30 .. 0x40]) rows01 = rows01._mm256_mullo_epi16(b: quants01) rows23 = rows23._mm256_mullo_epi16(b: quants23) rows45 = rows45._mm256_mullo_epi16(b: quants45) rows67 = rows67._mm256_mullo_epi16(b: quants67) // Permute: // // rows04 = i16x16 [a0 b0 .. g0 h0 // a4 b4 .. g4 h4] // rows31 = i16x16 [a3 b3 .. g3 e3 // a1 b1 .. g1 h1] // rows26 = i16x16 [a2 b2 .. g2 h2 // a6 b6 .. g6 h6] // rows75 = i16x16 [a7 b7 .. g7 h7 // a5 b5 .. g5 h5] rows04 = rows01._mm256_permute2x128_si256(b: rows45, imm8: 0x20) rows31 = rows23._mm256_permute2x128_si256(b: rows01, imm8: 0x31) rows26 = rows23._mm256_permute2x128_si256(b: rows67, imm8: 0x20) rows75 = rows67._mm256_permute2x128_si256(b: rows45, imm8: 0x31) // ---- // First pass (even rows). // This non-SIMD code: // // cb2 = (bq2 ~mod* 0x0000_29CF) ~mod+ (bq6 ~mod* 0x0000_1151) // cb6 = (bq2 ~mod* 0x0000_1151) ~mod+ (bq6 ~mod* 0xFFFF_D630) // // becomes: // // fp_rows62 = i16x16 [a6 b6 .. g6 h6 // a2 b2 .. g2 h2] // fp_bq2662ad = i16x16 [a2 a6 .. d2 d6 // a6 a2 .. d6 d2] // fp_bq2662eh = i16x16 [e2 e6 .. h2 h6 // e6 e2 .. h6 h2] // fp_cb26ad = i32x8 [cb2.a .. cb2.d // cb6.a .. cb6.d] // = i32x8 [a2*0x29CF+a6*0x1151 .. d2*0x29CF+d6*0x1151 // a6*0xD630+a2*0x1151 .. d6*0xD630+d2*0x1151] // fp_cb26eh = i32x8 [cb2.e .. cb2.h // cb6.e .. cb6.h] // = i32x8 [e2*0x29CF+e6*0x1151 .. h2*0x29CF+h6*0x1151 // e6*0xD630+e2*0x1151 .. h6*0xD630+h2*0x1151] fp_rows62 = rows26._mm256_permute2x128_si256(b: rows26, imm8: 0x01) fp_bq2662ad = rows26._mm256_unpacklo_epi16(b: fp_rows62) fp_bq2662eh = rows26._mm256_unpackhi_epi16(b: fp_rows62) fp_cb26ad = fp_bq2662ad._mm256_madd_epi16(b: k_29CF_1151_D630_1151) fp_cb26eh = fp_bq2662eh._mm256_madd_epi16(b: k_29CF_1151_D630_1151) // This non-SIMD code: // // ccp = (bq0 ~mod+ bq4) ~mod<< 13 // ccm = (bq0 ~mod- bq4) ~mod<< 13 // // becomes: // // fp_rows40pos = i16x16 [+a4 +b4 .. +g4 +h4 // +a0 +b0 .. +g0 +h0] // fp_rows04neg = i16x16 [+a0 +b0 .. +g0 +h0 // -a4 -b4 .. -g4 -h4] // fp_rows0pm4 = i16x16 [a0+a4 b0+b4 .. g0+g4 h0+h4 // a0-a4 b0-b4 .. g0-g4 h0-h4] // fp_ccpmad = i32x8 [ccp.a .. ccp.d // ccm.a .. ccm.d] // = i32x8 [(a0+a4)<<13 .. (d0+d4)<<13 // (a0-a4)<<13 .. (d0-d4)<<13] // fp_ccpmeh = i32x8 [ccp.e .. ccp.h // ccm.e .. ccm.h] // = i32x8 [(e0+e4)<<13 .. (h0+h4)<<13 // (e0-e4)<<13 .. (h0-h4)<<13] fp_rows40pos = rows04._mm256_permute2x128_si256(b: rows04, imm8: 0x01) fp_rows04neg = rows04._mm256_sign_epi16(b: k_0001_FFFF) fp_rows0pm4 = fp_rows40pos._mm256_add_epi16(b: fp_rows04neg) fp_ccpmad = k_0000._mm256_unpacklo_epi16(b: fp_rows0pm4)._mm256_srai_epi32(imm8: 16 - 13) fp_ccpmeh = k_0000._mm256_unpackhi_epi16(b: fp_rows0pm4)._mm256_srai_epi32(imm8: 16 - 13) // This non-SIMD code: // // cd0 = ccp ~mod+ cb2 // cd1 = ccm ~mod+ cb6 // cd2 = ccm ~mod- cb6 // cd3 = ccp ~mod- cb2 // // becomes: // // fp_cd01ad = i32x8 [cd0.a .. cd0.d // cd1.a .. cd1.d] // = i32x8 [ccp.a+cb2.a .. ccp.d+cb2.d // ccm.a+cb6.a .. ccm.d+cb6.d] // fp_cd01eh = i32x8 [cd0.e .. cd0.h // cd1.e .. cd1.h] // = i32x8 [ccp.e+cb2.e .. ccp.h+cb2.h // ccm.e+cb6.e .. ccm.h+cb6.h] // fp_cd32ad = i32x8 [cd3.a .. cd3.d // cd2.a .. cd2.d] // = i32x8 [ccp.a-cb2.a .. ccp.d-cb2.d // ccm.a-cb6.a .. ccm.d-cb6.d] // fp_cd32eh = i32x8 [cd3.e .. cd3.h // cd2.e .. cd2.h] // = i32x8 [ccp.e-cb2.e .. ccp.h-cb2.h // ccm.e-cb6.e .. ccm.h-cb6.h] fp_cd01ad = fp_ccpmad._mm256_add_epi32(b: fp_cb26ad) fp_cd01eh = fp_ccpmeh._mm256_add_epi32(b: fp_cb26eh) fp_cd32ad = fp_ccpmad._mm256_sub_epi32(b: fp_cb26ad) fp_cd32eh = fp_ccpmeh._mm256_sub_epi32(b: fp_cb26eh) // ---- // First pass (odd rows). // This non-SIMD code: // // ci73 = bq7 ~mod+ bq3 // ci51 = bq5 ~mod+ bq1 // cl73 = (ci73 ~mod* 0xFFFF_E6DC) ~mod+ (ci51 ~mod* 0x0000_25A1) // cl51 = (ci73 ~mod* 0x0000_25A1) ~mod+ (ci51 ~mod* 0x0000_1925) // // becomes: // // fp_sums7351 = i16x16 [a7+a3 b7+b3 .. g7+g3 h7+h3 // a5+a1 b5+b1 .. g5+g1 h5+h1] // fp_sums5173 = i16x16 [a5+a1 b5+b1 .. g5+g1 h5+h1 // a7+a3 b7+b3 .. g7+g3 h7+h3] // fp_ci73515173ad = i16x16 [a7+a3 a5+a1 .. d7+d3 d5+d1 // a5+a1 a7+a3 .. d5+d1 d7+d3] // fp_ci73515173eh = i16x16 [e7+e3 e5+e1 .. h7+h3 h5+d1 // e5+e1 e7+e3 .. h5+h1 h7+d3] // fp_cl7351ad = i32x8 [cl73.a .. cl73.d // cl51.a .. cl51.d] // = i32x8 [(a7+a3)*0xE6DC+(a5+a1)*0x25A1 .. (d7+d3)*0xE6DC+(d5+d1)*0x25A1 // (a5+a1)*0x1925+(a7+a3)*0x25A1 .. (d5+d1)*0x1925+(d7+d3)*0x25A1] // fp_cl7351eh = i32x8 [cl73.e .. cl73.h // cl51.e .. cl51.h] // = i32x8 [(e7+e3)*0xE6DC+(e5+e1)*0x25A1 .. (h7+h3)*0xE6DC+(h5+h1)*0x25A1 // (e5+e1)*0x1925+(e7+e3)*0x25A1 .. (h5+h1)*0x1925+(h7+h3)*0x25A1] fp_sums7351 = rows75._mm256_add_epi16(b: rows31) fp_sums5173 = fp_sums7351._mm256_permute2x128_si256(b: fp_sums7351, imm8: 0x01) fp_ci73515173ad = fp_sums7351._mm256_unpacklo_epi16(b: fp_sums5173) fp_ci73515173eh = fp_sums7351._mm256_unpackhi_epi16(b: fp_sums5173) fp_cl7351ad = fp_ci73515173ad._mm256_madd_epi16(b: k_E6DC_25A1_1925_25A1) fp_cl7351eh = fp_ci73515173eh._mm256_madd_epi16(b: k_E6DC_25A1_1925_25A1) // This non-SIMD code: // // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFF_ADFD) ~mod+ (bq5 ~mod* 0xFFFF_EFB0)) // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFF_E333) ~mod+ (bq7 ~mod* 0xFFFF_ECC1)) // // becomes: // // fp_rows13 = i16x16 [a1 b1 .. g1 e1 a3 b3 .. g3 h3] // fp_bq7153ad = i16x16 [a7 a1 .. d7 d1 a5 a3 .. d5 d3] // fp_bq7153eh = i16x16 [e7 e1 .. h7 h1 e5 e3 .. h5 h3] // fp_ck75ad = i32x8 [ck7.a .. ck7.d // ck5.a .. ck5.d] // = i32x8 [cl73.a+a7*0xECC1+a1*0xE333 .. cl73.d+d7*0xECC1+d1*0xE333 // cl51.a+a5*0xEFB0+a3*0xADFD .. cl51.d+d5*0xEFB0+d3*0xADFD] // fp_ck75eh = i32x8 [ck7.e .. ck7.h // ck5.e .. ck5.h] // = i32x8 [cl73.e+e7*0xECC1+e1*0xE333 .. cl73.h+h7*0xECC1+h1*0xE333 // cl51.e+e5*0xEFB0+e3*0xADFD .. cl51.h+h5*0xEFB0+h3*0xADFD] fp_rows13 = rows31._mm256_permute2x128_si256(b: rows31, imm8: 0x01) fp_bq7153ad = rows75._mm256_unpacklo_epi16(b: fp_rows13) fp_bq7153eh = rows75._mm256_unpackhi_epi16(b: fp_rows13) fp_ck75ad = fp_bq7153ad._mm256_madd_epi16(b: k_ECC1_E333_EFB0_ADFD)._mm256_add_epi32(b: fp_cl7351ad) fp_ck75eh = fp_bq7153eh._mm256_madd_epi16(b: k_ECC1_E333_EFB0_ADFD)._mm256_add_epi32(b: fp_cl7351eh) // This non-SIMD code: // // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x0000_133E) ~mod+ (bq7 ~mod* 0xFFFF_E333)) // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x0000_1051) ~mod+ (bq5 ~mod* 0xFFFF_ADFD)) // // becomes: // // fp_cl5173ad = i32x8 [cl51.a .. cl51.d // cl73.a .. cl73.d] // fp_cl5173eh = i32x8 [cl51.e .. cl51.h // cl73.e .. cl73.h] // fp_ck13ad = i32x8 [ck1.a .. ck1.d // ck3.a .. ck3.d] // = i32x8 [cl51.a+a7*0xE333+a1*0x133E .. cl51.d+d7*0xE333+d1*0x133E // cl73.a+a5*0xADFD+a3*0x1051 .. cl73.d+d5*0xADFD+d3*0x1051] // fp_ck13eh = i32x8 [ck1.e .. ck3.h // ck3.e .. ck1.h] // = i32x8 [cl51.e+e7*0xE333+e1*0x133E .. cl51.h+h7*0xE333+h1*0x133E // cl73.e+e5*0xADFD+e3*0x1051 .. cl73.h+h5*0xADFD+h3*0x1051] fp_cl5173ad = fp_cl7351ad._mm256_permute2x128_si256(b: fp_cl7351ad, imm8: 0x01) fp_cl5173eh = fp_cl7351eh._mm256_permute2x128_si256(b: fp_cl7351eh, imm8: 0x01) fp_ck13ad = fp_cl5173ad._mm256_add_epi32(b: fp_bq7153ad._mm256_madd_epi16(b: k_E333_133E_ADFD_1051)) fp_ck13eh = fp_cl5173eh._mm256_add_epi32(b: fp_bq7153eh._mm256_madd_epi16(b: k_E333_133E_ADFD_1051)) // ---- // First pass (combine rows). // We have now calculated: // // fp_ck13ad = i32x8 [ck1.a .. ck1.d ck3.a .. ck3.d] // fp_ck13eh = i32x8 [ck1.e .. ck1.h ck3.e .. ck3.h] // fp_ck75ad = i32x8 [ck7.a .. ck7.d ck5.a .. ck5.d] // fp_ck75eh = i32x8 [ck7.e .. ck7.h ck5.e .. ck5.h] // fp_cd01ad = i32x8 [cd0.a .. cd0.d cd1.a .. cd1.d] // fp_cd01eh = i32x8 [cd0.e .. cd0.h cd1.e .. cd1.h] // fp_cd32ad = i32x8 [cd3.a .. cd3.d cd2.a .. cd2.d] // fp_cd32eh = i32x8 [cd3.e .. cd3.h cd2.e .. cd2.h] // // This non-SIMD code: // // intermediate0 = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) // intermediate1 = this.util.sign_extend_rshift_u32(a: (cd1 ~mod+ ck3) ~mod+ (1 << 10), n: 11) // intermediate2 = this.util.sign_extend_rshift_u32(a: (cd2 ~mod+ ck5) ~mod+ (1 << 10), n: 11) // intermediate3 = this.util.sign_extend_rshift_u32(a: (cd3 ~mod+ ck7) ~mod+ (1 << 10), n: 11) // intermediate4 = this.util.sign_extend_rshift_u32(a: (cd3 ~mod- ck7) ~mod+ (1 << 10), n: 11) // intermediate5 = this.util.sign_extend_rshift_u32(a: (cd2 ~mod- ck5) ~mod+ (1 << 10), n: 11) // intermediate6 = this.util.sign_extend_rshift_u32(a: (cd1 ~mod- ck3) ~mod+ (1 << 10), n: 11) // intermediate7 = this.util.sign_extend_rshift_u32(a: (cd0 ~mod- ck1) ~mod+ (1 << 10), n: 11) // // becomes a mix of adds, subtracts, shifts and packs. intermediate01ad = fp_cd01ad._mm256_add_epi32(b: fp_ck13ad)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate01eh = fp_cd01eh._mm256_add_epi32(b: fp_ck13eh)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate01 = intermediate01ad._mm256_packs_epi32(b: intermediate01eh) intermediate32ad = fp_cd32ad._mm256_add_epi32(b: fp_ck75ad)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate32eh = fp_cd32eh._mm256_add_epi32(b: fp_ck75eh)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate32 = intermediate32ad._mm256_packs_epi32(b: intermediate32eh) intermediate45ad = fp_cd32ad._mm256_sub_epi32(b: fp_ck75ad)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate45eh = fp_cd32eh._mm256_sub_epi32(b: fp_ck75eh)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate45 = intermediate45ad._mm256_packs_epi32(b: intermediate45eh) intermediate76ad = fp_cd01ad._mm256_sub_epi32(b: fp_ck13ad)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate76eh = fp_cd01eh._mm256_sub_epi32(b: fp_ck13eh)._mm256_add_epi32(b: k_0400_0000)._mm256_srai_epi32(imm8: 11) intermediate76 = intermediate76ad._mm256_packs_epi32(b: intermediate76eh) // ---- // Transpose/re-order the intermediate rows and columns. We start with: // // intermediate01 = i16x16 [in.a0 .. in.d0 in.e0 .. in.h0 in.a1 .. in.d1 in.e1 .. in.h1] // intermediate32 = i16x16 [in.a3 .. in.d3 in.e3 .. in.h3 in.a2 .. in.d2 in.e2 .. in.h2] // intermediate45 = i16x16 [in.a4 .. in.d4 in.e4 .. in.h4 in.a5 .. in.d5 in.e5 .. in.h5] // intermediate76 = i16x16 [in.a7 .. in.d7 in.e7 .. in.h7 in.a6 .. in.d6 in.e6 .. in.h6] ita0a1e0e1 = intermediate01._mm256_permute4x64_epi64(imm8: 0xD8) ita2a3e2e3 = intermediate32._mm256_permute4x64_epi64(imm8: 0x72) ita4a5e4e5 = intermediate45._mm256_permute4x64_epi64(imm8: 0xD8) ita6a7e6e7 = intermediate76._mm256_permute4x64_epi64(imm8: 0x72) // We now have: // // ita0a1e0e1 = i16x16 [in.a0 .. in.d0 in.a1 .. in.d1 in.e0 .. in.h0 in.e1 .. in.h1] // ita2a3e2e3 = i16x16 [in.a2 .. in.d2 in.a3 .. in.d3 in.e2 .. in.h2 in.e3 .. in.h3] // ita4a5e4e5 = i16x16 [in.a4 .. in.d4 in.a5 .. in.d5 in.e4 .. in.h4 in.e5 .. in.h5] // ita6a7e6e7 = i16x16 [in.a6 .. in.d6 in.a7 .. in.d7 in.e6 .. in.h6 in.e7 .. in.h7] ita0c0e0g0 = ita0a1e0e1._mm256_unpacklo_epi16(b: ita2a3e2e3) ita1c1e1g1 = ita0a1e0e1._mm256_unpackhi_epi16(b: ita2a3e2e3) ita4c4e4g4 = ita4a5e4e5._mm256_unpacklo_epi16(b: ita6a7e6e7) ita5c5e5g5 = ita4a5e4e5._mm256_unpackhi_epi16(b: ita6a7e6e7) // We now have: // // ita0c0e0g0 = i16x16 [in.a0 in.a2 .. in.d0 in.d2 in.e0 in.e2 .. in.h0 in.h2] // ita1c1e1g1 = i16x16 [in.a1 in.a3 .. in.d1 in.d3 in.e1 in.e3 .. in.h1 in.h3] // ita4c4e4g4 = i16x16 [in.a4 in.a6 .. in.d4 in.d6 in.e4 in.e6 .. in.h4 in.h6] // ita5c5e5g5 = i16x16 [in.a5 in.a7 .. in.d5 in.d7 in.e5 in.e7 .. in.h5 in.h7] ita0b0e0f0 = ita0c0e0g0._mm256_unpacklo_epi16(b: ita1c1e1g1) itc0d0g0h0 = ita0c0e0g0._mm256_unpackhi_epi16(b: ita1c1e1g1) ita4b4e4f4 = ita4c4e4g4._mm256_unpacklo_epi16(b: ita5c5e5g5) itc4d4g4h4 = ita4c4e4g4._mm256_unpackhi_epi16(b: ita5c5e5g5) // We now have: // // ita0b0e0f0 = i16x16 [in.a0 .. in.a3 in.b0 .. in.b3 in.e0 .. in.e3 in.f0 .. in.f3] // itc0d0g0h0 = i16x16 [in.c0 .. in.c3 in.d0 .. in.d3 in.g0 .. in.g3 in.h0 .. in.h3] // ita4b4e4f4 = i16x16 [in.a4 .. in.a7 in.b4 .. in.b7 in.e4 .. in.e7 in.f4 .. in.f7] // itc4d4g4h4 = i16x16 [in.c4 .. in.c7 in.d4 .. in.d7 in.g4 .. in.g7 in.h4 .. in.h7] intermediateae = ita0b0e0f0._mm256_unpacklo_epi64(b: ita4b4e4f4) intermediatebf = ita0b0e0f0._mm256_unpackhi_epi64(b: ita4b4e4f4) intermediatecg = itc0d0g0h0._mm256_unpacklo_epi64(b: itc4d4g4h4) intermediatedh = itc0d0g0h0._mm256_unpackhi_epi64(b: itc4d4g4h4) // We now have: // // intermediateae = i16x16 [in.a0 .. in.a3 in.a4 .. in.a7 in.e0 .. in.e3 in.e4 .. in.e7] // intermediatebf = i16x16 [in.b0 .. in.b3 in.b4 .. in.b7 in.f0 .. in.f4 in.f5 .. in.f7] // intermediatecg = i16x16 [in.c0 .. in.c3 in.c4 .. in.c7 in.g0 .. in.g3 in.g4 .. in.g7] // intermediatedh = i16x16 [in.d0 .. in.d3 in.d4 .. in.d7 in.h0 .. in.h4 in.h5 .. in.h7] break.goto_do_second_pass }}.goto_do_second_pass // ---- // To recap, we have: // // intermediateae = i16x16 [in.a0 in.a1 .. in.a6 in.a7 // in.e0 in.e1 .. in.e6 in.e7] // intermediatebf = i16x16 [in.b0 in.b1 .. in.b6 in.b7 // in.f0 in.f1 .. in.f6 in.f7] // intermediatecg = i16x16 [in.c0 in.c1 .. in.c6 in.c7 // in.g0 in.g1 .. in.g6 in.g7] // intermediatedh = i16x16 [in.d0 in.d1 .. in.d6 in.d7 // in.h0 in.h1 .. in.h6 in.h7] // // Which is similar to this, but note the transposition and re-ordering: // // rows01 = i16x16 [a0 b0 .. g0 h0 // a1 b1 .. g1 h1] // rows23 = i16x16 [a2 b2 .. g2 e2 // a3 b3 .. g3 h3] // rows45 = i16x16 [a4 b4 .. g4 h4 // a5 b5 .. g5 h5] // rows67 = i16x16 [a6 b6 .. g6 h6 // a7 b7 .. g7 h7] // Permute: // // intermediatedb = i16x16 [in.d0 in.d1 .. in.d6 in.d7 // in.b0 in.b1 .. in.b6 in.b7] // intermediatehf = i16x16 [in.h0 in.h1 .. in.h6 in.h7 // in.f0 in.f1 .. in.f6 in.f7] // // The row/column analogies are: // // [a0 .. h0 a4 .. h4] = rows04 <=> intermediateae = [in.a0 .. in.a7 in.e0 .. in.e7] // [a3 .. h3 a1 .. h1] = rows31 <=> intermediatedb = [in.d0 .. in.d7 in.b0 .. in.b7] // [a2 .. h2 a6 .. h6] = rows26 <=> intermediatecg = [in.c0 .. in.c7 in.g0 .. in.g7] // [a7 .. h7 a5 .. h5] = rows75 <=> intermediatehf = [in.h0 .. in.h7 in.f0 .. in.f7] intermediatedb = intermediatedh._mm256_permute2x128_si256(b: intermediatebf, imm8: 0x20) intermediatehf = intermediatedh._mm256_permute2x128_si256(b: intermediatebf, imm8: 0x31) // ---- // Second pass (even columns). // This non-SIMD code: // // rb2 = (in2 ~mod* 0x0000_29CF) ~mod+ (in6 ~mod* 0x0000_1151) // rb6 = (in2 ~mod* 0x0000_1151) ~mod+ (in6 ~mod* 0xFFFF_D630) // // becomes: // // sp_rb26ad = i32x8 [rb2.a .. rb2.d rb6.a .. rb6.d] transposed // sp_rb26eh = i32x8 [rb2.e .. rb2.h rb6.e .. rb6.h] transposed // // 'Tranposed' means that the rows and column names should be swapped. // rb2.a, rb6.h, etc should really be rbc.0, rbg.7, etc but the former // notation lines up closer with the first pass' comments. sp_cols62 = intermediatecg._mm256_permute2x128_si256(b: intermediatecg, imm8: 0x01) sp_bq2662ad = intermediatecg._mm256_unpacklo_epi16(b: sp_cols62) sp_bq2662eh = intermediatecg._mm256_unpackhi_epi16(b: sp_cols62) sp_rb26ad = sp_bq2662ad._mm256_madd_epi16(b: k_29CF_1151_D630_1151) sp_rb26eh = sp_bq2662eh._mm256_madd_epi16(b: k_29CF_1151_D630_1151) // This non-SIMD code: // // rcp = (in0 ~mod+ in4) ~mod<< 13 // rcm = (in0 ~mod- in4) ~mod<< 13 // // becomes: // // sp_rcpmad = i32x8 [rcp.a .. rcp.d rcm.a .. rcm.d] transposed // sp_rcpmeh = i32x8 [rcp.e .. rcp.h rcm.e .. rcm.h] transposed sp_cols40pos = intermediateae._mm256_permute2x128_si256(b: intermediateae, imm8: 0x01) sp_cols04neg = intermediateae._mm256_sign_epi16(b: k_0001_FFFF) sp_cols0pm4 = sp_cols40pos._mm256_add_epi16(b: sp_cols04neg) sp_rcpmad = k_0000._mm256_unpacklo_epi16(b: sp_cols0pm4)._mm256_srai_epi32(imm8: 16 - 13) sp_rcpmeh = k_0000._mm256_unpackhi_epi16(b: sp_cols0pm4)._mm256_srai_epi32(imm8: 16 - 13) // This non-SIMD code: // // rd0 = rcp ~mod+ rb2 // rd1 = rcm ~mod+ rb6 // rd2 = rcm ~mod- rb6 // rd3 = rcp ~mod- rb2 // // becomes: // // sp_rd01ad = i32x8 [rd0.a .. rd0.d rd1.a .. rd1.d] transposed // sp_rd01eh = i32x8 [rd0.e .. rd0.h rd1.e .. rd1.h] transposed // sp_rd32ad = i32x8 [rd3.a .. rd3.d rd2.a .. rd2.d] transposed // sp_rd32eh = i32x8 [rd3.e .. rd3.h rd2.e .. rd2.h] transposed sp_rd01ad = sp_rcpmad._mm256_add_epi32(b: sp_rb26ad) sp_rd01eh = sp_rcpmeh._mm256_add_epi32(b: sp_rb26eh) sp_rd32ad = sp_rcpmad._mm256_sub_epi32(b: sp_rb26ad) sp_rd32eh = sp_rcpmeh._mm256_sub_epi32(b: sp_rb26eh) // ---- // Second pass (odd columns). // This non-SIMD code: // // ri73 = in7 ~mod+ in3 // ri51 = in5 ~mod+ in1 // rl73 = (ri73 ~mod* 0xFFFF_E6DC) ~mod+ (ri51 ~mod* 0x0000_25A1) // rl51 = (ri73 ~mod* 0x0000_25A1) ~mod+ (ri51 ~mod* 0x0000_1925) // // becomes: // // sp_rl7351ad = i32x8 [rl73.a .. rl73.d rl51.a .. rl51.d] transposed // sp_rl7351eh = i32x8 [rl73.e .. rl73.h rl51.e .. rl51.h] transposed sp_sums7351 = intermediatehf._mm256_add_epi16(b: intermediatedb) sp_sums5173 = sp_sums7351._mm256_permute2x128_si256(b: sp_sums7351, imm8: 0x01) sp_ri73515173ad = sp_sums7351._mm256_unpacklo_epi16(b: sp_sums5173) sp_ri73515173eh = sp_sums7351._mm256_unpackhi_epi16(b: sp_sums5173) sp_rl7351ad = sp_ri73515173ad._mm256_madd_epi16(b: k_E6DC_25A1_1925_25A1) sp_rl7351eh = sp_ri73515173eh._mm256_madd_epi16(b: k_E6DC_25A1_1925_25A1) // This non-SIMD code: // // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFF_ADFD) ~mod+ (in5 ~mod* 0xFFFF_EFB0)) // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFF_E333) ~mod+ (in7 ~mod* 0xFFFF_ECC1)) // // becomes: // // sp_rk75ad = i32x8 [rk7.a .. rk7.d rk5.a .. rk5.d] transposed // sp_rk75eh = i32x8 [rk7.e .. rk7.h rk5.e .. rk5.h] transposed sp_cols13 = intermediatedb._mm256_permute2x128_si256(b: intermediatedb, imm8: 0x01) sp_bq7153ad = intermediatehf._mm256_unpacklo_epi16(b: sp_cols13) sp_bq7153eh = intermediatehf._mm256_unpackhi_epi16(b: sp_cols13) sp_rk75ad = sp_bq7153ad._mm256_madd_epi16(b: k_ECC1_E333_EFB0_ADFD)._mm256_add_epi32(b: sp_rl7351ad) sp_rk75eh = sp_bq7153eh._mm256_madd_epi16(b: k_ECC1_E333_EFB0_ADFD)._mm256_add_epi32(b: sp_rl7351eh) // This non-SIMD code: // // rk1 = rl51 ~mod+ ((in1 ~mod* 0x0000_133E) ~mod+ (in7 ~mod* 0xFFFF_E333)) // rk3 = rl73 ~mod+ ((in3 ~mod* 0x0000_1051) ~mod+ (in5 ~mod* 0xFFFF_ADFD)) // // becomes: // // sp_rk13ad = i32x8 [rk1.a .. rk1.d rk3.a .. rk3.d] transposed // sp_rk13eh = i32x8 [rk1.e .. rk3.h rk3.e .. rk1.h] transposed sp_rl5173ad = sp_rl7351ad._mm256_permute2x128_si256(b: sp_rl7351ad, imm8: 0x01) sp_rl5173eh = sp_rl7351eh._mm256_permute2x128_si256(b: sp_rl7351eh, imm8: 0x01) sp_rk13ad = sp_rl5173ad._mm256_add_epi32(b: sp_bq7153ad._mm256_madd_epi16(b: k_E333_133E_ADFD_1051)) sp_rk13eh = sp_rl5173eh._mm256_add_epi32(b: sp_bq7153eh._mm256_madd_epi16(b: k_E333_133E_ADFD_1051)) // ---- // Second pass (combine columns). // We have now calculated: // // sp_rk13ad = i32x8 [rk1.a .. rk1.d rk3.a .. rk3.d] transposed // sp_rk13eh = i32x8 [rk1.e .. rk1.h rk3.e .. rk3.h] transposed // sp_rk75ad = i32x8 [rk7.a .. rk7.d rk5.a .. rk5.d] transposed // sp_rk75eh = i32x8 [rk7.e .. rk7.h rk5.e .. rk5.h] transposed // sp_rd01ad = i32x8 [rd0.a .. rd0.d rd1.a .. rd1.d] transposed // sp_rd01eh = i32x8 [rd0.e .. rd0.h rd1.e .. rd1.h] transposed // sp_rd32ad = i32x8 [rd3.a .. rd3.d rd2.a .. rd2.d] transposed // sp_rd32eh = i32x8 [rd3.e .. rd3.h rd2.e .. rd2.h] transposed // // This non-SIMD code: // // final0 = this.util.sign_extend_rshift_u32(a: (rd0 ~mod+ rk1) ~mod+ (1 << 17), n: 18) // final1 = this.util.sign_extend_rshift_u32(a: (rd1 ~mod+ rk3) ~mod+ (1 << 17), n: 18) // final2 = this.util.sign_extend_rshift_u32(a: (rd2 ~mod+ rk5) ~mod+ (1 << 17), n: 18) // final3 = this.util.sign_extend_rshift_u32(a: (rd3 ~mod+ rk7) ~mod+ (1 << 17), n: 18) // final4 = this.util.sign_extend_rshift_u32(a: (rd3 ~mod- rk7) ~mod+ (1 << 17), n: 18) // final5 = this.util.sign_extend_rshift_u32(a: (rd2 ~mod- rk5) ~mod+ (1 << 17), n: 18) // final6 = this.util.sign_extend_rshift_u32(a: (rd1 ~mod- rk3) ~mod+ (1 << 17), n: 18) // final7 = this.util.sign_extend_rshift_u32(a: (rd0 ~mod- rk1) ~mod+ (1 << 17), n: 18) // // becomes a mix of adds, subtracts, shifts and packs. final01ad = sp_rd01ad._mm256_add_epi32(b: sp_rk13ad)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final01eh = sp_rd01eh._mm256_add_epi32(b: sp_rk13eh)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final01 = final01ad._mm256_packs_epi32(b: final01eh) final32ad = sp_rd32ad._mm256_add_epi32(b: sp_rk75ad)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final32eh = sp_rd32eh._mm256_add_epi32(b: sp_rk75eh)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final32 = final32ad._mm256_packs_epi32(b: final32eh) final45ad = sp_rd32ad._mm256_sub_epi32(b: sp_rk75ad)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final45eh = sp_rd32eh._mm256_sub_epi32(b: sp_rk75eh)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final45 = final45ad._mm256_packs_epi32(b: final45eh) final76ad = sp_rd01ad._mm256_sub_epi32(b: sp_rk13ad)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final76eh = sp_rd01eh._mm256_sub_epi32(b: sp_rk13eh)._mm256_add_epi32(b: k_0000_0002)._mm256_srai_epi32(imm8: 18) final76 = final76ad._mm256_packs_epi32(b: final76eh) // ---- // Transpose/re-order the final rows and columns. fta0a1e0e1 = final01._mm256_permute4x64_epi64(imm8: 0xD8) fta2a3e2e3 = final32._mm256_permute4x64_epi64(imm8: 0x72) fta4a5e4e5 = final45._mm256_permute4x64_epi64(imm8: 0xD8) fta6a7e6e7 = final76._mm256_permute4x64_epi64(imm8: 0x72) fta0c0e0g0 = fta0a1e0e1._mm256_unpacklo_epi16(b: fta2a3e2e3) fta1c1e1g1 = fta0a1e0e1._mm256_unpackhi_epi16(b: fta2a3e2e3) fta4c4e4g4 = fta4a5e4e5._mm256_unpacklo_epi16(b: fta6a7e6e7) fta5c5e5g5 = fta4a5e4e5._mm256_unpackhi_epi16(b: fta6a7e6e7) fta0b0e0f0 = fta0c0e0g0._mm256_unpacklo_epi16(b: fta1c1e1g1) ftc0d0g0h0 = fta0c0e0g0._mm256_unpackhi_epi16(b: fta1c1e1g1) fta4b4e4f4 = fta4c4e4g4._mm256_unpacklo_epi16(b: fta5c5e5g5) ftc4d4g4h4 = fta4c4e4g4._mm256_unpackhi_epi16(b: fta5c5e5g5) finalae = fta0b0e0f0._mm256_unpacklo_epi64(b: fta4b4e4f4) finalbf = fta0b0e0f0._mm256_unpackhi_epi64(b: fta4b4e4f4) finalcg = ftc0d0g0h0._mm256_unpacklo_epi64(b: ftc4d4g4h4) finaldh = ftc0d0g0h0._mm256_unpackhi_epi64(b: ftc4d4g4h4) // ---- // Pack i16x16 into i8x32 (using signed saturation; i8 ranges from -128 ..= // +127), bias by +0x80 (u8 ranges from 0 ..= 255) and extract the rows. // // We also drop the "'Tranposed' means that the rows and column names // should be swapped". Rows are back to 0 ..= 7 and columns are a ..= h. final0145 = finalae._mm256_packs_epi16(b: finalbf)._mm256_add_epi8(b: k_8080) final2367 = finalcg._mm256_packs_epi16(b: finaldh)._mm256_add_epi8(b: k_8080) final0 = final0145._mm256_extract_epi64(index: 0) final1 = final0145._mm256_extract_epi64(index: 1) final2 = final2367._mm256_extract_epi64(index: 0) final3 = final2367._mm256_extract_epi64(index: 1) final4 = final0145._mm256_extract_epi64(index: 2) final5 = final0145._mm256_extract_epi64(index: 3) final6 = final2367._mm256_extract_epi64(index: 2) final7 = final2367._mm256_extract_epi64(index: 3) // ---- // Write to the args.dst_buffer. if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final0) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final1) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final2) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final3) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final4) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final5) args.dst_buffer = remaining if args.dst_stride > args.dst_buffer.length() { return nothing } assert 8 <= args.dst_buffer.length() via "a <= b: a <= c; c <= b"(c: args.dst_stride) assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() remaining = args.dst_buffer[args.dst_stride ..] args.dst_buffer.poke_u64le!(a: final6) args.dst_buffer = remaining if 8 > args.dst_buffer.length() { return nothing } assert args.dst_buffer.length() >= 8 via "a >= b: b <= a"() args.dst_buffer.poke_u64le!(a: final7) }