fix bugs in dot

diff --git a/source/NE10_add.neon.s b/source/NE10_add.neon.s
index 19d09e6..2111348 100644 (file)
--- a/source/NE10_add.neon.s
+++ b/source/NE10_add.neon.s
@@ -120,7 +120,7 @@ add_vec2f_neon:
         @  r3: int count & the number of items in the input array that can be
         @                   processed in chunks of 4 vectors
         @
-        @  r4:  the number of items that are residual that will be processed at the begin of
+        @  r4: the number of items that are residual that will be processed at the begin of
         @                   the input array
         @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 

diff --git a/source/NE10_dot.neon.s b/source/NE10_dot.neon.s
index 8452361..0f03f3d 100644 (file)
--- a/source/NE10_dot.neon.s
+++ b/source/NE10_dot.neon.s
@@ -26,7 +26,7 @@
 
 
 
-        .balign   4
+        .align   4
         .global   dot_vec2f_neon
         .thumb
         .thumb_func
@@ -45,80 +45,55 @@ dot_vec2f_neon:
         @  r3: int count & the number of items in the input array that can be
         @                   processed in chunks of 4 vectors
         @
-        @  r4:  the number of items that are left to be processed at the end of
+        @  r4: the number of items that are residual that will be processed at the begin of
         @                   the input array
         @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
         push              {r4}
-        and               r4, r3, #3          @ r4 = count % 4;
-        sub               r3, r3, r4          @ count = count - r3; This is what's left to be processed after this loop
+        and               r4, r3, #3          @ r4 = count % 4; calculate the residual loop
+        asr               r3, r3, #2          @ r3 = count >> 2; calculate the main loop
 
-        cbz               r3, .L_check_vec2
+        cbz               r4, .L_check_mainloop_vec2
 
-        @ load the 1st set of values
-          vld2.32         {q0-q1}, [r1]!
-          vld2.32         {q2-q3}, [r2]!
-          subs            r3, r3, #8          @ 4 for this set, and 4 for the 2nd set
+.L_residualloop_vec2:
+        @ process the residual items in the input array
+        vld1.f32          d0, [r1]!
+        vld1.f32          d1, [r2]!
 
-        @ calculate values for the 1st set
-          vmul.f32        q8, q0, q2
-          vmla.f32        q8, q1, q3
+        subs              r4, r4, #1
 
-        @ load the 2nd set of values
-          vld2.32         {q0-q1}, [r1]!
-          vld2.32         {q2-q3}, [r2]!
+        @ calculate values
+        vmul.f32          d0, d0, d1
+        vpadd.f32         d0, d0
 
-          ble             .L_mainloopend_vec2
+        vst1.32           {d0[0]}, [r0]!
 
-.L_mainloop_vec2:
-        @ store the result for the 1st/next (e.g. 3rd) set
-          vst1.32         {d16,d17}, [r0]!
+        bgt               .L_residualloop_vec2
 
-        @ calculate values for the 2nd/next (e.g. 3rd) set
-          vmul.f32        q8, q0, q2
-          vmla.f32        q8, q1, q3
+.L_check_mainloop_vec2:
+        cbz               r3, .L_return_vec2
 
-       @ load the next (e.g. 3rd) set of values
+        @ load the current set of values
           vld2.32         {q0-q1}, [r1]!
           vld2.32         {q2-q3}, [r2]!
-          subs            r3, r3, #4
-
-        bgt             .L_mainloop_vec2             @ loop if r2 is > r3, if we have at least another 4 vectors (8 floats) to process
 
-.L_mainloopend_vec2:
-        @ the last iteration for this call
-        @ store the result for the set of values before the last one (e.g 2nd set)
-          vst1.32         {d16,d17}, [r0]!
-
-        @ calculate values for the last (e.g. 3rd) set
-          vmul.f32        q8, q0, q2
-          vmla.f32        q8, q1, q3
-
-        @ store the result for the last (e.g. 3rd) set
-          vst1.32         {d16,d17}, [r0]!
-
-.L_check_vec2:
-     @ check if anything left to process at the end of the input array
-        cmp               r4, #0
-        ble               .L_return_vec2
-
-.L_secondloop_vec2:
-     @ process the last few items left in the input array
-        vld1.f32          d0, [r1]!
-        vld1.f32          d1, [r2]!
-
-        subs              r4, r4, #1
+.L_mainloop_vec2:
+        @ calculate values for current set
+        vmul.f32        q8, q0, q2
+        vmla.f32        q8, q1, q3
 
-        @ calculate values
-        vmul.f32          d0, d0, d1
-        vpadd.f32         d0, d0
+        @ store the result for current set
+        vst1.32         {d16,d17}, [r0]!
+        subs            r3, r3, #1
 
-        vst1.32           {d0[0]}, [r0]!
+        @ load the next set of values
+        vld2.32         {q0-q1}, [r1]!
+        vld2.32         {q2-q3}, [r2]!
 
-        bgt               .L_secondloop_vec2
+        bgt             .L_mainloop_vec2             @ loop if r3 > 0, if we have at least another 4 vectors (8 floats) to process
 
 .L_return_vec2:
-     @ return
+        @ return
         pop               {r4}
         mov               r0, #0
         bx                lr
@@ -126,7 +101,7 @@ dot_vec2f_neon:
 
 
 
-        .align  2
+        .align  4
         .global dot_vec3f_neon
         .thumb
         .thumb_func
@@ -144,75 +119,18 @@ dot_vec3f_neon:
         @  r3: int count & the number of items in the input array that can be
         @                   processed in chunks of 4 vectors
         @
-        @  r4:  the number of items that are left to be processed at the end of
+        @  r4: the number of items that are residual that will be processed at the begin of
         @                   the input array
         @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
         push              {r4}
-        and               r4, r3, #3          @ r3 = count % 4;
-        sub               r3, r3, r4          @ count = count - r3; This is what's left to be processed after this loop
+        and               r4, r3, #3          @ r4 = count % 4; calculate the residual loop
+        asr               r3, r3, #2          @ r3 = count >> 2; calculate the main loop
 
-        cmp               r3, #0
-        beq               .L_check_vec3
+        cbz               r4, .L_check_mainloop_vec3
 
-        @ load the 1st set of values
-          vld3.32         {d0, d2, d4}, [r1]!
-          vld3.32         {d1, d3, d5}, [r1]!
-          vld3.32         {d16, d18, d20}, [r2]!
-          vld3.32         {d17, d19, d21}, [r2]!
-          subs            r3, r3, #8          @ 4 for this set, and 4 for the 2nd set
-
-        @ calculate values for the 1st set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-
-        @ load the 2nd set of values
-          vld3.32         {d0, d2, d4}, [r1]!
-          vld3.32         {d1, d3, d5}, [r1]!
-          vld3.32         {d16, d18, d20}, [r2]!
-          vld3.32         {d17, d19, d21}, [r2]!
-
-          ble             .L_mainloopend_vec3
-
-.L_mainloop_vec3:
-        @ store the result for the 1st/next (e.g. 3rd) set
-          vst1.32         {d30, d31}, [r0]!
-
-        @ calculate values for the 2nd/next (e.g. 3rd) set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-
-       @ load the next (e.g. 3rd) set of values
-          vld3.32         {d0, d2, d4}, [r1]!
-          vld3.32         {d1, d3, d5}, [r1]!
-          vld3.32         {d16, d18, d20}, [r2]!
-          vld3.32         {d17, d19, d21}, [r2]!
-          subs            r3, r3, #4
-
-        bgt               .L_mainloop_vec3             @ loop if r2 is > r3, if we have at least another 4 vectors (12 floats) to process
-
-.L_mainloopend_vec3:
-        @ the last iteration for this call
-        @ store the result for the set of values before the last one (e.g 2nd set)
-          vst1.32         {d30, d31}, [r0]!
-
-        @ calculate values for the last (e.g. 3rd) set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-
-        @ store the result for the last (e.g. 3rd) set
-          vst1.32         {d30, d31}, [r0]!
-
-.L_check_vec3:
-     @ check if anything left to process at the end of the input array
-        cmp               r4, #0
-        ble               .L_return_vec3
-
-.L_secondloop_vec3:
-     @ process the last few items left in the input array
+.L_residualloop_vec3:
+        @ process the residual items in the input array
         vld3.f32          {d0[0], d2[0], d4[0]}, [r1]!     @ The values are loaded like so:
                                                            @      q0 = { V1.x, -, -, - };
                                                            @      q1 = { V1.y, -, -, - };
@@ -231,10 +149,37 @@ dot_vec3f_neon:
 
         vst1.32           {d0[0]}, [r0]!
 
-        bgt               .L_secondloop_vec3
+        bgt               .L_residualloop_vec3
+
+.L_check_mainloop_vec3:
+        cbz               r3, .L_return_vec3
+
+        @ load current set of values
+        vld3.32         {d0, d2, d4}, [r1]!
+        vld3.32         {d1, d3, d5}, [r1]!
+        vld3.32         {d16, d18, d20}, [r2]!
+        vld3.32         {d17, d19, d21}, [r2]!
+
+.L_mainloop_vec3:
+        @ calculate values for current set
+        vmul.f32        q15, q0, q8
+        vmla.f32        q15, q1, q9
+        vmla.f32        q15, q2, q10
+
+        @ store the result for current set
+        vst1.32         {d30, d31}, [r0]!
+        subs            r3, r3, #1
+
+        @ load the next set of values
+        vld3.32         {d0, d2, d4}, [r1]!
+        vld3.32         {d1, d3, d5}, [r1]!
+        vld3.32         {d16, d18, d20}, [r2]!
+        vld3.32         {d17, d19, d21}, [r2]!
+
+        bgt               .L_mainloop_vec3             @ loop if r3 > 0, if we have at least another 4 vectors (12 floats) to process
 
 .L_return_vec3:
-     @ return
+        @ return
         pop               {r4}
         mov               r0, #0
         bx                lr
@@ -242,7 +187,7 @@ dot_vec3f_neon:
 
 
 
-        .align  2
+        .align  4
         .global dot_vec4f_neon
         .thumb
         .thumb_func
@@ -260,104 +205,65 @@ dot_vec4f_neon:
         @  r3: int count & the number of items in the input array that can be
         @                   processed in chunks of 4 vectors
         @
-        @  r4:  the number of items that are left to be processed at the end of
+        @  r4: the number of items that are residual that will be processed at the begin of
         @                   the input array
         @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
         push              {r4}
-        and               r4, r3, #3          @ r4 = count % 4;
-        sub               r3, r3, r4          @ count = count - r4; This is what's left to be processed after this loop
+        and               r4, r3, #3          @ r4 = count % 4; calculate the residual loop
+        asr               r3, r3, #2          @ r3 = count >> 2; calculate the main loop
+
+        cbz               r4, .L_check_mainloop_vec4
 
-        cmp               r3, #0
-        beq               .L_check_vec4
+.L_residualloop_vec4:
+        @ process the residual items in the input array
+        vld1.f32          {d0, d1}, [r1]!     @ The values are loaded like so:
+                                                                  @      q0 = { V1.x, V1.y, V1.z, V1.w };
+        vld1.f32          {d2, d3}, [r2]!     @ The values are loaded like so:
+                                                                  @      q1 = { V2.x, V2.y, V2.z, V2.w };
 
-        @ load the 1st set of values
-          vld4.32         {d0, d2, d4, d6}, [r1]!
-          vld4.32         {d1, d3, d5, d7}, [r1]!
-          vld4.32         {d16, d18, d20, d22}, [r2]!
-          vld4.32         {d17, d19, d21, d23}, [r2]!
+        subs              r4, r4, #1
 
-          subs            r3, r3, #8          @ 4 for this set, and 4 for the 2nd set
+        @ calculate values
+        vmul.f32          q0, q0, q1
+        vadd.f32          d0, d0, d1
+        vpadd.f32         d0, d0
 
-        @ calculate values for the 1st set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-          vmla.f32        q15, q3, q11
+        vst1.32          {d0[0]}, [r0]!
 
-        @ load the 2nd set of values
-          vld4.32         {d0, d2, d4, d6}, [r1]!
-          vld4.32         {d1, d3, d5, d7}, [r1]!
-          vld4.32         {d16, d18, d20, d22}, [r2]!
-          vld4.32         {d17, d19, d21, d23}, [r2]!
+        bgt               .L_residualloop_vec4
 
-          ble             .L_mainloopend_vec4
+.L_check_mainloop_vec4:
+        cbz               r3, .L_return_vec4
+
+        @ load current set of values
+        vld4.32         {d0, d2, d4, d6}, [r1]!
+        vld4.32         {d1, d3, d5, d7}, [r1]!
+        vld4.32         {d16, d18, d20, d22}, [r2]!
+        vld4.32         {d17, d19, d21, d23}, [r2]!
 
 .L_mainloop_vec4:
-        @ store the result for the 1st/next (e.g. 3rd) set
-          vst1.32         {d30, d31}, [r0]!
-
-        @ calculate values for the 2nd/next (e.g. 3rd) set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-          vmla.f32        q15, q3, q11
-
-       @ load the next (e.g. 3rd) set of values
-          vld4.32         {d0, d2, d4, d6}, [r1]!
-          vld4.32         {d1, d3, d5, d7}, [r1]!
-          vld4.32         {d16, d18, d20, d22}, [r2]!
-          vld4.32         {d17, d19, d21, d23}, [r2]!
-          subs            r3, r3, #4
-
-        bgt               .L_mainloop_vec4             @ loop if r2 is > r3, if we have at least another 4 vectors (16 floats) to process
-
-.L_mainloopend_vec4:
-        @ the last iteration for this call
-        @ store the result for the set of values before the last one (e.g 2nd set)
-          vst1.32         {d30, d31}, [r0]!
-
-        @ calculate values for the last (e.g. 3rd) set
-          vmul.f32        q15, q0, q8
-          vmla.f32        q15, q1, q9
-          vmla.f32        q15, q2, q10
-          vmla.f32        q15, q3, q11
-
-        @ store the result for the last (e.g. 3rd) set
-          vst1.32         {d30, d31}, [r0]!
-
-.L_check_vec4:
-     @ check if anything left to process at the end of the input array
-        cmp               r4, #0
-        ble               .L_return_vec4
-
-.L_secondloop_vec4:
-     @ process the last few items left in the input array
-        vld4.f32          {d0[0], d2[0], d4[0], d6[0]}, [r1]!     @ The values are loaded like so:
-                                                                  @      q0 = { V1.x, -, -, - };
-                                                                  @      q1 = { V1.y, -, -, - };
-                                                                  @      q2 = { V1.z, -, -, - };
-                                                                  @      q3 = { V1.w, -, -, - };
-        vld4.f32          {d1[0], d3[0], d5[0], d7[0]}, [r2]!     @ The values are loaded like so:
-                                                                  @      q0 = { V1.x, -, V2.x, - };
-                                                                  @      q1 = { V1.y, -, V2.y, - };
-                                                                  @      q2 = { V1.z, -, V2.z, - };
-                                                                  @      q3 = { V1.w, -, V2.w, - };
+        @ calculate values for current set
+        vmul.f32        q15, q0, q8
+        vmla.f32        q15, q1, q9
+        vmla.f32        q15, q2, q10
+        vmla.f32        q15, q3, q11
 
-        subs              r4, r4, #1
+        @ store the result for current set
+        vst1.32         {d30, d31}, [r0]!
+        subs            r3, r3, #1
 
-        @ calculate values
-        vmul.f32          d0, d0, d1
-        vmla.f32          d0, d2, d3
-        vmla.f32          d0, d4, d5
-        vmla.f32          d0, d6, d7
+        @ load the next set of values
+        vld4.32         {d0, d2, d4, d6}, [r1]!
+        vld4.32         {d1, d3, d5, d7}, [r1]!
+        vld4.32         {d16, d18, d20, d22}, [r2]!
+        vld4.32         {d17, d19, d21, d23}, [r2]!
 
-        vst1.32          {d0[0]}, [r0]!
-
-        bgt               .L_secondloop_vec4
+        bgt               .L_mainloop_vec4             @ loop if r3 > 0, if we have at least another 4 vectors (12 floats) to process
 
 .L_return_vec4:
-     @ return
+        @ return
         pop               {r4}
         mov               r0, #0
         bx                lr
+