SIMD: remove outdated SSE2 code
authorErik de Castro Lopo <erikd@mega-nerd.com>
Tue, 21 Feb 2017 09:27:05 +0000 (20:27 +1100)
committerErik de Castro Lopo <erikd@mega-nerd.com>
Tue, 21 Feb 2017 09:27:07 +0000 (20:27 +1100)
Removes FLAC__lpc_restore_signal_16_intrin_sse2() which was faster
than than C code, but not faster than MMX-accelerated ASM functions.
It's also slower than the new SSE4.1 functions that were added by
the previous patch.

Patch-from: lvqcl <lvqcl.mail@gmail.com>

src/libFLAC/include/private/lpc.h
src/libFLAC/lpc_intrin_sse2.c
src/libFLAC/stream_decoder.c

index 14f2ef3..63d6432 100644 (file)
@@ -206,9 +206,6 @@ void FLAC__lpc_restore_signal_wide_asm_ia32(const FLAC__int32 residual[], uint32
 #    endif /* FLAC__HAS_NASM */
 #  endif /* FLAC__CPU_IA32 */
 #  if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN
-#    ifdef FLAC__SSE2_SUPPORTED
-void FLAC__lpc_restore_signal_16_intrin_sse2(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[]);
-#    endif
 #    ifdef FLAC__SSE4_1_SUPPORTED
 void FLAC__lpc_restore_signal_intrin_sse41(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[]);
 void FLAC__lpc_restore_signal_16_intrin_sse41(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[]);
index cb97b9d..d750138 100644 (file)
@@ -931,157 +931,6 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse2(const FLAC__in
        }
 }
 
-#if defined FLAC__CPU_IA32 && !defined FLAC__HAS_NASM /* unused for x64; not better than MMX asm */
-
-#define DATA16_RESULT(xmmN) curr = *residual++ + (_mm_cvtsi128_si32(xmmN) >> lp_quantization); *data++ = curr;
-
-FLAC__SSE_TARGET("sse2")
-void FLAC__lpc_restore_signal_16_intrin_sse2(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[])
-{
-       if (order < 8 || order > 12) {
-               FLAC__lpc_restore_signal(residual, data_len, qlp_coeff, order, lp_quantization, data);
-               return;
-       }
-       if (data_len == 0)
-               return;
-
-       FLAC__ASSERT(order >= 8);
-       FLAC__ASSERT(order <= 12);
-
-       if(order > 8) { /* order == 9, 10, 11, 12 */
-               FLAC__int32 curr;
-               __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
-               xmm0 = _mm_loadu_si128((const __m128i*)(qlp_coeff+0));
-               xmm6 = _mm_loadu_si128((const __m128i*)(qlp_coeff+4));
-               xmm1 = _mm_loadu_si128((const __m128i*)(qlp_coeff+8)); /* read 0 to 3 uninitialized coeffs... */
-               switch(order)                                          /* ...and zero them out */
-               {
-               case 9:
-                       xmm1 = _mm_slli_si128(xmm1, 12); xmm1 = _mm_srli_si128(xmm1, 12); break;
-               case 10:
-                       xmm1 = _mm_slli_si128(xmm1, 8); xmm1 = _mm_srli_si128(xmm1, 8); break;
-               case 11:
-                       xmm1 = _mm_slli_si128(xmm1, 4); xmm1 = _mm_srli_si128(xmm1, 4); break;
-               }
-               xmm2 = _mm_setzero_si128();
-               xmm0 = _mm_packs_epi32(xmm0, xmm6);
-               xmm1 = _mm_packs_epi32(xmm1, xmm2);
-
-               xmm4 = _mm_loadu_si128((const __m128i*)(data-12));
-               xmm5 = _mm_loadu_si128((const __m128i*)(data-8));
-               xmm3 = _mm_loadu_si128((const __m128i*)(data-4));
-               xmm4 = _mm_shuffle_epi32(xmm4, _MM_SHUFFLE(0,1,2,3));
-               xmm5 = _mm_shuffle_epi32(xmm5, _MM_SHUFFLE(0,1,2,3));
-               xmm3 = _mm_shuffle_epi32(xmm3, _MM_SHUFFLE(0,1,2,3));
-               xmm4 = _mm_packs_epi32(xmm4, xmm2);
-               xmm3 = _mm_packs_epi32(xmm3, xmm5);
-
-               xmm7 = _mm_slli_si128(xmm1, 2);
-               xmm7 = _mm_or_si128(xmm7, _mm_srli_si128(xmm0, 14));
-               xmm2 = _mm_slli_si128(xmm0, 2);
-
-               /* xmm0, xmm1: qlp_coeff
-                  xmm2, xmm7: qlp_coeff << 16 bit
-                  xmm3, xmm4: data */
-
-               xmm5 = _mm_madd_epi16(xmm4, xmm1);
-               xmm6 = _mm_madd_epi16(xmm3, xmm0);
-               xmm6 = _mm_add_epi32(xmm6, xmm5);
-               xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-               xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-               DATA16_RESULT(xmm6);
-
-               data_len--;
-
-               if(data_len % 2) {
-                       xmm6 = _mm_srli_si128(xmm3, 14);
-                       xmm4 = _mm_slli_si128(xmm4, 2);
-                       xmm3 = _mm_slli_si128(xmm3, 2);
-                       xmm4 = _mm_or_si128(xmm4, xmm6);
-                       xmm3 = _mm_insert_epi16(xmm3, curr, 0);
-
-                       xmm5 = _mm_madd_epi16(xmm4, xmm1);
-                       xmm6 = _mm_madd_epi16(xmm3, xmm0);
-                       xmm6 = _mm_add_epi32(xmm6, xmm5);
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-                       DATA16_RESULT(xmm6);
-
-                       data_len--;
-               }
-
-               while(data_len) { /* data_len is a multiple of 2 */
-                       /* 1 _mm_slli_si128 per data element less but we need shifted qlp_coeff in xmm2:xmm7 */
-                       xmm6 = _mm_srli_si128(xmm3, 12);
-                       xmm4 = _mm_slli_si128(xmm4, 4);
-                       xmm3 = _mm_slli_si128(xmm3, 4);
-                       xmm4 = _mm_or_si128(xmm4, xmm6);
-                       xmm3 = _mm_insert_epi16(xmm3, curr, 1);
-
-                       xmm5 = _mm_madd_epi16(xmm4, xmm7);
-                       xmm6 = _mm_madd_epi16(xmm3, xmm2);
-                       xmm6 = _mm_add_epi32(xmm6, xmm5);
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-                       DATA16_RESULT(xmm6);
-
-                       xmm3 = _mm_insert_epi16(xmm3, curr, 0);
-
-                       xmm5 = _mm_madd_epi16(xmm4, xmm1);
-                       xmm6 = _mm_madd_epi16(xmm3, xmm0);
-                       xmm6 = _mm_add_epi32(xmm6, xmm5);
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-                       DATA16_RESULT(xmm6);
-
-                       data_len-=2;
-               }
-       } /* endif(order > 8) */
-       else {
-               FLAC__int32 curr;
-               __m128i xmm0, xmm1, xmm3, xmm6;
-               xmm0 = _mm_loadu_si128((const __m128i*)(qlp_coeff+0));
-               xmm1 = _mm_loadu_si128((const __m128i*)(qlp_coeff+4));
-               xmm0 = _mm_packs_epi32(xmm0, xmm1);
-
-               xmm1 = _mm_loadu_si128((const __m128i*)(data-8));
-               xmm3 = _mm_loadu_si128((const __m128i*)(data-4));
-               xmm1 = _mm_shuffle_epi32(xmm1, _MM_SHUFFLE(0,1,2,3));
-               xmm3 = _mm_shuffle_epi32(xmm3, _MM_SHUFFLE(0,1,2,3));
-               xmm3 = _mm_packs_epi32(xmm3, xmm1);
-
-               /* xmm0: qlp_coeff
-                  xmm3: data */
-
-               xmm6 = _mm_madd_epi16(xmm3, xmm0);
-               xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-               xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-               DATA16_RESULT(xmm6);
-
-               data_len--;
-
-               while(data_len) {
-                       xmm3 = _mm_slli_si128(xmm3, 2);
-                       xmm3 = _mm_insert_epi16(xmm3, curr, 0);
-
-                       xmm6 = _mm_madd_epi16(xmm3, xmm0);
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 8));
-                       xmm6 = _mm_add_epi32(xmm6, _mm_srli_si128(xmm6, 4));
-
-                       DATA16_RESULT(xmm6);
-
-                       data_len--;
-               }
-       }
-}
-
-#endif /* defined FLAC__CPU_IA32 && !defined FLAC__HAS_NASM */
-
 #endif /* FLAC__SSE2_SUPPORTED */
 #endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */
 #endif /* FLAC__NO_ASM */
index 4c62d23..509a6ab 100644 (file)
@@ -394,11 +394,6 @@ static FLAC__StreamDecoderInitStatus init_stream_internal_(
                }
 #endif
 #if FLAC__HAS_X86INTRIN && ! defined FLAC__INTEGER_ONLY_LIBRARY
-# if defined FLAC__SSE2_SUPPORTED && !defined FLAC__HAS_NASM /* OPT_SSE: not better than MMX asm */
-               if(decoder->private_->cpuinfo.ia32.sse2) {
-                       decoder->private_->local_lpc_restore_signal_16bit = FLAC__lpc_restore_signal_16_intrin_sse2;
-               }
-# endif
 # if defined FLAC__SSE4_1_SUPPORTED
                if(decoder->private_->cpuinfo.ia32.sse41) {
                        decoder->private_->local_lpc_restore_signal = FLAC__lpc_restore_signal_intrin_sse41;