SIMD-accelerated integer convsamp routine for MIPS DSPr2

git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/trunk@1059 632fc199-4ca6-4c93-a231-07263d6284db

diff --git a/simd/jsimd.h b/simd/jsimd.h
index e5c5d44006978c1c9d9478124d1065c24ce802b5..f75aa9bbaa141c560d1ac93c64d61e2f69eb511b 100644 (file)
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -651,6 +651,10 @@ EXTERN(void) jsimd_convsamp_neon JPP((JSAMPARRAY sample_data,
                                       JDIMENSION start_col,
                                       DCTELEM * workspace));
 
+EXTERN(void) jsimd_convsamp_mips_dspr2 JPP((JSAMPARRAY sample_data,
+                                            JDIMENSION start_col,
+                                            DCTELEM * workspace));
+
 EXTERN(void) jsimd_convsamp_float_3dnow JPP((JSAMPARRAY sample_data,
                                              JDIMENSION start_col,
                                              FAST_FLOAT * workspace));

diff --git a/simd/jsimd_mips.c b/simd/jsimd_mips.c
index d8d6b199a942bf65ff9a0eb9acddb615829dfea5..5b4f89334246bb571623fd3bd65de64833fef4e7 100644 (file)
--- a/simd/jsimd_mips.c
+++ b/simd/jsimd_mips.c
@@ -453,6 +453,21 @@ jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
 GLOBAL(int)
 jsimd_can_convsamp (void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_MIPS_DSPR2)
+    return 1;
+
   return 0;
 }
 
@@ -483,6 +498,8 @@ GLOBAL(void)
 jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
                 DCTELEM * workspace)
 {
+  if (simd_support & JSIMD_MIPS_DSPR2)
+    jsimd_convsamp_mips_dspr2(sample_data, start_col, workspace);
 }
 
 GLOBAL(void)

diff --git a/simd/jsimd_mips_dspr2.S b/simd/jsimd_mips_dspr2.S
index d478a6d20c19f7a6950c042c2095d4861d9c6fbd..198c3496b73b8fcec833e71993bb9a0b0c65de13 100644 (file)
--- a/simd/jsimd_mips_dspr2.S
+++ b/simd/jsimd_mips_dspr2.S
@@ -2812,6 +2812,155 @@ LEAF_MIPS_DSPR2(jsimd_idct_12x12_pass2_mips_dspr2)
 
 END(jsimd_idct_12x12_pass2_mips_dspr2)
 
+/*****************************************************************************/
+LEAF_MIPS_DSPR2(jsimd_convsamp_mips_dspr2)
+/*
+ * a0     - sample_data
+ * a1     - start_col
+ * a2     - workspace
+ */
+
+    lw             t0, 0(a0)
+    li             t7, 0xff80ff80
+    addu           t0, t0, a1
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    lw             t0, 4(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 0(a2)
+    usw            t4, 4(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 8(a2)
+    usw            t6, 12(a2)
+
+    lw             t0, 8(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 16(a2)
+    usw            t4, 20(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 24(a2)
+    usw            t6, 28(a2)
+
+    lw             t0, 12(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 32(a2)
+    usw            t4, 36(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 40(a2)
+    usw            t6, 44(a2)
+
+    lw             t0, 16(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 48(a2)
+    usw            t4, 52(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 56(a2)
+    usw            t6, 60(a2)
+
+    lw             t0, 20(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 64(a2)
+    usw            t4, 68(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 72(a2)
+    usw            t6, 76(a2)
+
+    lw             t0, 24(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 80(a2)
+    usw            t4, 84(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 88(a2)
+    usw            t6, 92(a2)
+
+    lw             t0, 28(a0)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu           t0, t0, a1
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    ulw            t1, 0(t0)
+    ulw            t2, 4(t0)
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 96(a2)
+    usw            t4, 100(a2)
+    preceu.ph.qbr  t3, t1
+    preceu.ph.qbl  t4, t1
+    usw            t5, 104(a2)
+    usw            t6, 108(a2)
+    preceu.ph.qbr  t5, t2
+    preceu.ph.qbl  t6, t2
+    addu.ph        t3, t3, t7
+    addu.ph        t4, t4, t7
+    addu.ph        t5, t5, t7
+    addu.ph        t6, t6, t7
+    usw            t3, 112(a2)
+    usw            t4, 116(a2)
+    usw            t5, 120(a2)
+    usw            t6, 124(a2)
+
+    j              ra
+     nop
+
+END(jsimd_convsamp_mips_dspr2)
+
 /*****************************************************************************/
 LEAF_MIPS_DSPR2(jsimd_convsamp_float_mips_dspr2)
 /*