diff options
| -rw-r--r-- | apps/codecs/libatrac/atrac3_arm.S | 150 |
1 files changed, 48 insertions, 102 deletions
diff --git a/apps/codecs/libatrac/atrac3_arm.S b/apps/codecs/libatrac/atrac3_arm.S index 0dacff0b7c..68f8de2c4e 100644 --- a/apps/codecs/libatrac/atrac3_arm.S +++ b/apps/codecs/libatrac/atrac3_arm.S @@ -93,6 +93,47 @@ atrac3_iqmf_matrixing: * } * Note: r12 is a scratch register and can be used without restorage. ****************************************************************************/ + +/* To be called as first block to call smull for initial filling of the result + * registers lr/r9 and r12/r8. */ +#define DEWIN_8_SAMPLES_MUL_ASM \ + ldmia r2!, {r4, r5}; /* load win[0..1] */ \ + ldmia r1!, {r6, r7}; /* load in [0..1] */ \ + smull lr , r9, r4, r6; /* s1 = win[0] * in[0] */ \ + smull r12, r8, r5, r7; /* s2 = win[1] * in[1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ + +/* Called after first block. Will always multiply-add to the result registers + * lr/r9 and r12/r8. */ +#define DEWIN_8_SAMPLES_MLA_ASM \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ \ + ldmia r2!, {r4, r5}; /* load win[i..i+1] */ \ + ldmia r1!, {r6, r7}; /* load in [i..i+1] */ \ + smlal lr , r9, r4, r6; /* s1 = win[i ] * in[i ] */ \ + smlal r12, r8, r5, r7; /* s2 = win[i+1] * in[i+1] */ + .align 2 .global atrac3_iqmf_dewindowing .type atrac3_iqmf_dewindowing, %function @@ -105,108 +146,13 @@ atrac3_iqmf_dewindowing: stmfd sp!, {r4-r9, lr} /* save non-scratch registers */ .iqmf_dewindow_outer_loop: /* outer loop 0...counter-1 */ - /* 0.. 7 */ - ldmia r2!, {r4, r5} /* load win[0..1] */ - ldmia r1!, {r6, r7} /* load in[0..1] */ - smull lr , r9, r4, r6 /* s1 = win[0] * in[0] */ - smull r12, r8, r5, r7 /* s2 = win[1] * in[1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - /* 8..15 */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - /* 16..23 */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - /* 24..31 */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - /* 32..39 */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - /* 40..47 */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ - ldmia r2!, {r4, r5} /* load win[i...i+1] */ - ldmia r1!, {r6, r7} /* load in[i...i+1] */ - smlal lr , r9, r4, r6 /* s1 = win[i ] * in[i ] */ - smlal r12, r8, r5, r7 /* s2 = win[i+1] * in[i+1] */ + + DEWIN_8_SAMPLES_MUL_ASM /* 0.. 7, use "MUL" macro here! */ + DEWIN_8_SAMPLES_MLA_ASM /* 8..15 */ + DEWIN_8_SAMPLES_MLA_ASM /* 16..23 */ + DEWIN_8_SAMPLES_MLA_ASM /* 24..31 */ + DEWIN_8_SAMPLES_MLA_ASM /* 32..39 */ + DEWIN_8_SAMPLES_MLA_ASM /* 40..47 */ mov lr , lr , lsr #31 orr r9, lr , r9, lsl #1 /* s1 = low>>31 || hi<<1 */ |