audio-resampler: add more neon optimizations

diff --git a/gst-libs/gst/audio/audio-resampler-neon.h b/gst-libs/gst/audio/audio-resampler-neon.h
index 1785bbe..8643bc2 100644 (file)
--- a/gst-libs/gst/audio/audio-resampler-neon.h
+++ b/gst-libs/gst/audio/audio-resampler-neon.h
@@ -115,9 +115,6 @@ static inline void
 inner_product_gint16_cubic_1_neon (gint16 * o, const gint16 * a,
     const gint16 * b, gint len, const gint16 * icoeff)
 {
-    uint32_t remainder = len % 4;
-    len = len - remainder;
-
     asm volatile ("      vmov.s32 q0, #0\n"
                   "      cmp %[len], #0\n"
                   "      beq 2f\n"
@@ -133,15 +130,6 @@ inner_product_gint16_cubic_1_neon (gint16 * o, const gint16 * a,
                   "      bne 1b\n"
                   "      vadd.s32 q0, q0, q1\n"
                   "2:"
-                  "      cmp %[remainder], #0\n"
-                  "      beq 4f\n"
-                  "3:"
-                  "      vld1.16 {d16}, [%[b]]!\n"
-                  "      vld1.16 {d20[]}, [%[a]]!\n"
-                  "      subs %[remainder], %[remainder], #1\n"
-                  "      vmlal.s16 q0, d16, d20\n"
-                  "      bne 3b\n"
-                  "4:"
                   "      vshrn.s32 d0, q0, #15\n"
                   "      vld1.16 {d20}, [%[ic]]\n"
                   "      vmull.s16 q0, d0, d20\n"
@@ -150,7 +138,7 @@ inner_product_gint16_cubic_1_neon (gint16 * o, const gint16 * a,
                   "      vqrshrn.s32 d0, q0, #15\n"
                   "      vst1.s16 d0[0], [%[o]]\n"
                   : [a] "+r" (a), [b] "+r" (b),
-                    [len] "+r" (len), [remainder] "+r" (remainder)
+                    [len] "+r" (len)
                   : [o] "r" (o), [ic] "r" (icoeff)
                   : "cc", "q0", "q1",
                     "d16", "d17", "d18", "d19",
@@ -161,18 +149,118 @@ static inline void
 inner_product_gint32_none_1_neon (gint32 * o, const gint32 * a,
     const gint32 * b, gint len, const gint32 * icoeff)
 {
+    uint32_t remainder = len % 8;
+    len = len - remainder;
+
+    asm volatile ("      vmov.s64 q0, #0\n"
+                  "      cmp %[len], #0\n"
+                  "      beq 2f\n"
+                  "      vmov.s64 q1, #0\n"
+                  "1:"
+                  "      vld1.32 {d16, d17, d18, d19}, [%[b]]!\n"
+                  "      vld1.32 {d20, d21, d22, d23}, [%[a]]!\n"
+                  "      subs %[len], %[len], #8\n"
+                  "      vmlal.s32 q0, d16, d20\n"
+                  "      vmlal.s32 q1, d17, d21\n"
+                  "      vmlal.s32 q0, d18, d22\n"
+                  "      vmlal.s32 q1, d19, d23\n"
+                  "      bne 1b\n"
+                  "      vadd.s64 q0, q0, q1\n"
+                  "2:"
+                  "      cmp %[remainder], #0\n"
+                  "      beq 4f\n"
+                  "3:"
+                  "      vld1.32 {d16, d17}, [%[b]]!\n"
+                  "      vld1.32 {d20, d21}, [%[a]]!\n"
+                  "      subs %[remainder], %[remainder], #4\n"
+                  "      vmlal.s32 q0, d16, d20\n"
+                  "      vmlal.s32 q0, d17, d21\n"
+                  "      bne 3b\n"
+                  "4:"
+                  "      vadd.s64 d0, d0, d1\n"
+                  "      vqrshrn.s64 d0, q0, #31\n"
+                  "      vst1.s32 d0[0], [%[o]]\n"
+                  : [a] "+r" (a), [b] "+r" (b),
+                    [len] "+r" (len), [remainder] "+r" (remainder)
+                  : [o] "r" (o)
+                  : "cc", "q0", "q1",
+                    "d16", "d17", "d18", "d19",
+                    "d20", "d21", "d22", "d23");
 }
 
 static inline void
 inner_product_gint32_linear_1_neon (gint32 * o, const gint32 * a,
     const gint32 * b, gint len, const gint32 * icoeff)
 {
+    asm volatile ("      vmov.s64 q0, #0\n"
+                  "      cmp %[len], #0\n"
+                  "      beq 2f\n"
+                  "      vmov.s64 q1, #0\n"
+                  "1:"
+                  "      vld1.s32 {d16, d17, d18, d19}, [%[b]]!\n"
+                  "      vld2.s32 {d20[], d21[]}, [%[a]]!\n"
+                  "      vld2.s32 {d22[], d23[]}, [%[a]]!\n"
+                  "      subs %[len], %[len], #4\n"
+                  "      vmlal.s32 q0, d16, d20\n"
+                  "      vmlal.s32 q1, d17, d21\n"
+                  "      vmlal.s32 q0, d18, d22\n"
+                  "      vmlal.s32 q1, d19, d23\n"
+                  "      bne 1b\n"
+                  "      vadd.s64 q0, q0, q1\n"
+                  "2:"
+                  "      vld1.s32 {d20}, [%[ic]]\n"
+                  "      vshrn.s64 d0, q0, #31\n"
+                  "      vmull.s32 q0, d0, d20\n"
+                  "      vadd.s64 d0, d0, d1\n"
+                  "      vqrshrn.s64 d0, q0, #31\n"
+                  "      vst1.s32 d0[0], [%[o]]\n"
+                  : [a] "+r" (a), [b] "+r" (b),
+                    [len] "+r" (len)
+                  : [o] "r" (o), [ic] "r" (icoeff)
+                  : "cc", "q0", "q1",
+                    "d16", "d17", "d18", "d19",
+                    "d20", "d21", "d22", "d23", "memory");
 }
 
 static inline void
 inner_product_gint32_cubic_1_neon (gint32 * o, const gint32 * a,
     const gint32 * b, gint len, const gint32 * icoeff)
 {
+    asm volatile ("      vmov.s64 q0, #0\n"
+                  "      cmp %[len], #0\n"
+                  "      beq 2f\n"
+                  "      vmov.s64 q1, #0\n"
+                  "1:"
+                  "      vld1.s32 {q4, q5}, [%[b]]!\n"
+                  "      vld1.s32 {q6, q7}, [%[b]]!\n"
+                  "      vld1.s32 {d16[], d17[]}, [%[a]]!\n"
+                  "      vld1.s32 {d18[], d19[]}, [%[a]]!\n"
+                  "      vld1.s32 {d20[], d21[]}, [%[a]]!\n"
+                  "      vld1.s32 {d22[], d23[]}, [%[a]]!\n"
+                  "      subs %[len], %[len], #4\n"
+                  "      vmlal.s32 q0, d16, d8\n"
+                  "      vmlal.s32 q1, d17, d9\n"
+                  "      vmlal.s32 q0, d18, d10\n"
+                  "      vmlal.s32 q1, d19, d11\n"
+                  "      vmlal.s32 q0, d20, d12\n"
+                  "      vmlal.s32 q1, d21, d13\n"
+                  "      vmlal.s32 q0, d22, d14\n"
+                  "      vmlal.s32 q1, d23, d15\n"
+                  "      bne 1b\n"
+                  "2:"
+                  "      vld1.s32 {d20, d21}, [%[ic]]\n"
+                  "      vshrn.s64 d16, q0, #31\n"
+                  "      vshrn.s64 d17, q1, #31\n"
+                  "      vmull.s32 q0, d20, d16\n"
+                  "      vmlal.s32 q0, d21, d17\n"
+                  "      vadd.s64 d0, d0, d1\n"
+                  "      vqrshrn.s64 d0, q0, #31\n"
+                  "      vst1.s32 d0[0], [%[o]]\n"
+                  : [a] "+r" (a), [b] "+r" (b),
+                    [len] "+r" (len)
+                  : [o] "r" (o), [ic] "r" (icoeff)
+                  : "cc", "q0", "q1", "q4", "q5", "q6", "q7", "q8",
+                    "q9", "q10", "q11", "memory");
 }
 
 static inline void
@@ -216,7 +304,6 @@ inner_product_gfloat_none_1_neon (gfloat * o, const gfloat * a,
                   : [o] "r" (o)
                   : "cc", "q0", "q1", "q4", "q5", "q6", "q7", "q8",
                     "q9", "q10", "q11");
-
 }
 
 static inline void
@@ -247,7 +334,7 @@ inner_product_gfloat_linear_1_neon (gfloat * o, const gfloat * a,
                   "3:"
                   "      vld2.f32 {q4}, [%[b]]!\n"
                   "      vld1.f32 {q8}, [%[a]]!\n"
-                  "      subs %[remainder], %[remainder], #2\n"
+                  "      subs %[remainder], %[remainder], #4\n"
                   "      vmla.f32 q0, q4, q8\n"
                   "      bne 3b\n"
                   "4:"
@@ -267,9 +354,6 @@ static inline void
 inner_product_gfloat_cubic_1_neon (gfloat * o, const gfloat * a,
     const gfloat * b, gint len, const gfloat * icoeff)
 {
-    uint32_t remainder = len % 4;
-    len = len - remainder;
-
     asm volatile ("      vmov.f32 q0, #0.0\n"
                   "      cmp %[len], #0\n"
                   "      beq 2f\n"
@@ -289,22 +373,13 @@ inner_product_gfloat_cubic_1_neon (gfloat * o, const gfloat * a,
                   "      bne 1b\n"
                   "      vadd.f32 q0, q0, q1\n"
                   "2:"
-                  "      cmp %[remainder], #0\n"
-                  "      beq 4f\n"
-                  "3:"
-                  "      vld1.f32 {q4}, [%[b]]!\n"
-                  "      vld1.f32 {d16[], d17[]}, [%[a]]!\n"
-                  "      subs %[remainder], %[remainder], #1\n"
-                  "      vmla.f32 q0, q4, q8\n"
-                  "      bne 3b\n"
-                  "4:"
                   "      vld1.f32 {q10}, [%[ic]]\n"
                   "      vmul.f32 q0, q0, q10\n"
                   "      vadd.f32 d0, d0, d1\n"
                   "      vpadd.f32 d0, d0, d0\n"
                   "      vst1.f32 d0[0], [%[o]]\n"
                   : [a] "+r" (a), [b] "+r" (b),
-                    [len] "+r" (len), [remainder] "+r" (remainder)
+                    [len] "+r" (len)
                   : [o] "r" (o), [ic] "r" (icoeff)
                   : "cc", "q0", "q1", "q4", "q5", "q6", "q7", "q8",
                     "q9", "q10", "q11", "memory");
@@ -365,14 +440,15 @@ audio_resampler_check_neon (const gchar *target_name, const gchar *option)
     resample_gint16_linear_1 = resample_gint16_linear_1_neon;
     resample_gint16_cubic_1 = resample_gint16_cubic_1_neon;
 
+    resample_gint32_none_1 = resample_gint32_none_1_neon;
+    resample_gint32_linear_1 = resample_gint32_linear_1_neon;
+    resample_gint32_cubic_1 = resample_gint32_cubic_1_neon;
+
     resample_gfloat_none_1 = resample_gfloat_none_1_neon;
     resample_gfloat_linear_1 = resample_gfloat_linear_1_neon;
     resample_gfloat_cubic_1 = resample_gfloat_cubic_1_neon;
 
     if (0) {
-      resample_gint32_none_1 = resample_gint32_none_1_neon;
-      resample_gint32_linear_1 = resample_gint32_linear_1_neon;
-      resample_gint32_cubic_1 = resample_gint32_cubic_1_neon;
 
       resample_gdouble_none_1 = resample_gdouble_none_1_neon;
       resample_gdouble_linear_1 = resample_gdouble_linear_1_neon;