summaryrefslogtreecommitdiff
path: root/src/libFLAC
diff options
context:
space:
mode:
Diffstat (limited to 'src/libFLAC')
-rw-r--r--src/libFLAC/include/private/lpc.h3
-rw-r--r--src/libFLAC/lpc_intrin_sse2.c151
-rw-r--r--src/libFLAC/stream_decoder.c5
3 files changed, 0 insertions, 159 deletions
diff --git a/src/libFLAC/include/private/lpc.h b/src/libFLAC/include/private/lpc.h
index 14f2ef36..63d64324 100644
--- a/src/libFLAC/include/private/lpc.h
+++ b/src/libFLAC/include/private/lpc.h
@@ -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[]);
diff --git a/src/libFLAC/lpc_intrin_sse2.c b/src/libFLAC/lpc_intrin_sse2.c
index cb97b9de..d750138c 100644
--- a/src/libFLAC/lpc_intrin_sse2.c
+++ b/src/libFLAC/lpc_intrin_sse2.c
@@ -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 */
diff --git a/src/libFLAC/stream_decoder.c b/src/libFLAC/stream_decoder.c
index 4c62d232..509a6abd 100644
--- a/src/libFLAC/stream_decoder.c
+++ b/src/libFLAC/stream_decoder.c
@@ -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;
View Lorry Controller Status