Thom Johansen's first EMAC optimisation for the Coldfire - about a 3%-4% speedup

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@6024 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 222 insertions, 0 deletions

diff --git a/apps/codecs/libFLAC/SOURCES b/apps/codecs/libFLAC/SOURCES
index fc793f5..7f5abc2 100644
--- a/apps/codecs/libFLAC/SOURCES
+++ b/apps/codecs/libFLAC/SOURCES
@@ -10,3 +10,6 @@ md5.c
 memory.c
 seekable_stream_decoder.c
 stream_decoder.c
+#if CONFIG_CPU==MCF5249
+coldfire.c
+#endif
diff --git a/apps/codecs/libFLAC/coldfire.c b/apps/codecs/libFLAC/coldfire.c
new file mode 100644
index 0000000..c763c20
--- /dev/null
+++ b/apps/codecs/libFLAC/coldfire.c
@@ -0,0 +1,165 @@
+#ifndef SIMULATOR
+#include <private/coldfire.h>
+
+void FLAC__lpc_restore_signal_order8_mac(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
+{
+	register const FLAC__int32 *qlp0 = &qlp_coeff[(order-1)];
+	register FLAC__int32 sum;
+	register const FLAC__int32 *history;
+	
+	SET_MACSR(0);
+	history = &data[(-order)];
+	SET_ACC(0, acc0);
+	
+	switch (order) {
+	case 8:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 12(%2), %%d1, %%acc0\n\t"
+				"mov.l -12(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 16(%2), %%d1, %%acc0\n\t"
+				"mov.l -16(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 20(%2), %%d1, %%acc0\n\t"
+				"mov.l -20(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 24(%2), %%d1, %%acc0\n\t"
+				"mov.l -24(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 28(%2), %%d1, %%acc0\n\t"
+				"mov.l -28(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		}
+		return;
+	case 7:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 12(%2), %%d1, %%acc0\n\t"
+				"mov.l -12(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 16(%2), %%d1, %%acc0\n\t"
+				"mov.l -16(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 20(%2), %%d1, %%acc0\n\t"
+				"mov.l -20(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 24(%2), %%d1, %%acc0\n\t"
+				"mov.l -24(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		} 
+		return;
+	case 6:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 12(%2), %%d1, %%acc0\n\t"
+				"mov.l -12(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 16(%2), %%d1, %%acc0\n\t"
+				"mov.l -16(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 20(%2), %%d1, %%acc0\n\t"
+				"mov.l -20(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		} 
+		return;
+	case 5:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 12(%2), %%d1, %%acc0\n\t"
+				"mov.l -12(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 16(%2), %%d1, %%acc0\n\t"
+				"mov.l -16(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		} 
+		return;
+	case 4:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 12(%2), %%d1, %%acc0\n\t"
+				"mov.l -12(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		} 
+		return;
+	case 3:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, 8(%2), %%d1, %%acc0\n\t"
+				"mov.l -8(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		}
+		return;
+	case 2:
+		for( ; data_len != 0; --data_len) {
+			asm volatile(
+				"mov.l (%1), %%d0\n\t"
+				"mov.l (%2), %%d1\n\t"
+				"mac.l %%d0, %%d1, 4(%2), %%d1, %%acc0\n\t"
+				"mov.l -4(%1), %%d0\n\t"
+				"mac.l %%d0, %%d1, %%acc0\n\t"
+				"movclr.l %%acc0, %0"
+				: "=ad" (sum) : "a" (qlp0), "a" (history) : "d0", "d1");
+			++history;
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		}
+		return;
+	case 1:
+		// won't gain anything by using mac here. 
+		for( ; data_len != 0; --data_len) {
+			sum = (qlp0[0] * (*(history++)));
+			*(data++) = *(residual++) + (sum >> lp_quantization);
+		} 
+		return;
+	}
+}
+
+#endif
diff --git a/apps/codecs/libFLAC/include/private/coldfire.h b/apps/codecs/libFLAC/include/private/coldfire.h
new file mode 100644
index 0000000..22f1711
--- /dev/null
+++ b/apps/codecs/libFLAC/include/private/coldfire.h
@@ -0,0 +1,46 @@
+#ifndef SIMULATOR
+#ifndef _FLAC_COLDFIRE_H
+#define _FLAC_COLDFIRE_H
+
+#include <FLAC/ordinals.h>
+
+#define MACL(x, y, acc) \
+	asm volatile ("mac.l %0, %1, %%" #acc \
+	              : : "ad" ((x)), "ad" ((y)));
+
+#define MACL_SHIFT(x, y, shift, acc) \
+	asm volatile ("mac.l %0, %1, #" #shift ", %%" #acc \
+	              : : "ad" ((x)), "ad" ((y)));
+
+#define MSACL(x, y, acc) \
+	asm volatile ("msac.l %0, %1, %%" #acc \
+	    : : "ad" ((x)), "ad" ((y)));
+
+#define MSACL_SHIFT(x, y, shift, acc) \
+	asm volatile ("msac.l %0, %1, #" #shift ", %%" #acc \
+	              : : "ad" ((x)), "ad" ((y)));
+
+#define SET_MACSR(x) \
+	asm volatile ("mov.l %0, %%macsr" : : "adi" ((x)));
+
+#define TRANSFER_ACC(acca, accb) \
+	asm volatile ("mov.l %" #acca ", %" #accb);
+
+#define SET_ACC(x, acc) \
+	asm volatile ("mov.l %0, %%" #acc : : "adi" ((x)));
+
+#define GET_ACC(x, acc) \
+	asm volatile ("mov.l %%" #acc ", %0\n\t" : "=ad" ((x)));
+
+#define GET_ACC_CLR(x, acc) \
+	asm volatile ("movclr.l %%" #acc ", %0\n\t" : "=ad" ((x)));
+
+#define EMAC_SATURATE   0x00000080
+#define EMAC_FRACTIONAL 0x00000020
+#define EMAC_ROUND      0x00000010
+
+
+void FLAC__lpc_restore_signal_order8_mac(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[]);
+
+#endif
+#endif
diff --git a/apps/codecs/libFLAC/stream_decoder.c b/apps/codecs/libFLAC/stream_decoder.c
index ec43314..eb78d18 100644
--- a/apps/codecs/libFLAC/stream_decoder.c
+++ b/apps/codecs/libFLAC/stream_decoder.c
@@ -43,6 +43,10 @@
 #include "private/lpc.h"
 #include "private/memory.h"
 
+#if CONFIG_CPU==MCF5249
+#include <private/coldfire.h>
+#endif
+
 #ifdef HAVE_CONFIG_H
 #include <config.h>
 #endif
@@ -298,7 +302,11 @@ FLAC_API FLAC__StreamDecoderState FLAC__stream_decoder_init(FLAC__StreamDecoder
 	decoder->private_->local_lpc_restore_signal = FLAC__lpc_restore_signal;
 	decoder->private_->local_lpc_restore_signal_64bit = FLAC__lpc_restore_signal_wide;
 	decoder->private_->local_lpc_restore_signal_16bit = FLAC__lpc_restore_signal;
+#if CONFIG_CPU==MCF5249 && !SIMULATOR
+	decoder->private_->local_lpc_restore_signal_16bit_order8 = FLAC__lpc_restore_signal_order8_mac;
+#else
 	decoder->private_->local_lpc_restore_signal_16bit_order8 = FLAC__lpc_restore_signal;
+#endif
 	/* now override with asm where appropriate */
 #ifndef FLAC__NO_ASM
 	if(decoder->private_->cpuinfo.use_asm) {