Merge pull request #839 from Atcold/fix_ASIMD

Fix compilation for ASIMD, fix #766

3 files changed, 75 insertions, 237 deletions

diff --git a/lib/TH/CMakeLists.txt b/lib/TH/CMakeLists.txt
index 29343c7..e6cf91d 100644
--- a/lib/TH/CMakeLists.txt
+++ b/lib/TH/CMakeLists.txt
@@ -62,10 +62,13 @@ ENDIF (WITH_OPENMP)
 
 # ARM specific flags
 FIND_PACKAGE(ARM)
-IF (NEON_FOUND)
+IF (ASIMD_FOUND)
+  MESSAGE(STATUS "asimd/Neon found with compiler flag : -D__NEON__")
+  SET(CMAKE_C_FLAGS "-D__NEON__ ${CMAKE_C_FLAGS}")
+ELSEIF (NEON_FOUND)
   MESSAGE(STATUS "Neon found with compiler flag : -mfpu=neon -D__NEON__")
   SET(CMAKE_C_FLAGS "-mfpu=neon -D__NEON__ ${CMAKE_C_FLAGS}")
-ENDIF (NEON_FOUND)
+ENDIF (ASIMD_FOUND)
 IF (CORTEXA8_FOUND)
   MESSAGE(STATUS "Cortex-A8 Found with compiler flag : -mcpu=cortex-a8")
   SET(CMAKE_C_FLAGS "-mcpu=cortex-a8 -fprefetch-loop-arrays ${CMAKE_C_FLAGS}")
diff --git a/lib/TH/cmake/FindARM.cmake b/lib/TH/cmake/FindARM.cmake
index cf1f8fd..59c78d8 100644
--- a/lib/TH/cmake/FindARM.cmake
+++ b/lib/TH/cmake/FindARM.cmake
@@ -13,6 +13,15 @@ IF(CMAKE_SYSTEM_NAME MATCHES "Linux")
       set(NEON_FOUND false CACHE BOOL "NEON available on host")
    ENDIF (NEON_TRUE)
 
+   # on ARMv8, neon is inherit and instead listed as 'asimd' in /proc/cpuinfo
+   STRING(REGEX REPLACE "^.*(asimd).*$" "\\1" ASIMD_THERE ${CPUINFO})
+   STRING(COMPARE EQUAL "asimd" "${ASIMD_THERE}" ASIMD_TRUE)
+   IF (ASIMD_TRUE)
+      set(ASIMD_FOUND true CACHE BOOL "ASIMD/NEON available on host")
+   ELSE (ASIMD_TRUE)
+      set(ASIMD_FOUND false CACHE BOOL "ASIMD/NEON available on host")
+   ENDIF (ASIMD_TRUE)
+
    #Find the processor type (for now OMAP3 or OMAP4)
    STRING(REGEX REPLACE "^.*(OMAP3).*$" "\\1" OMAP3_THERE ${CPUINFO})
    STRING(COMPARE EQUAL "OMAP3" "${OMAP3_THERE}" OMAP3_TRUE)
diff --git a/lib/TH/vector/NEON.c b/lib/TH/vector/NEON.c
index bc7cb2b..327b006 100644
--- a/lib/TH/vector/NEON.c
+++ b/lib/TH/vector/NEON.c
@@ -1,252 +1,78 @@
 static void THFloatVector_fill_NEON(float *x, const float c, const ptrdiff_t n) {
-  float ctemp = c;
-  float * caddr = &ctemp;
-  __asm__ __volatile__ (
-      "mov         r0, %0           @ \n\t"
-      "ldr         r4, [%1]         @ \n\t"
-      "vdup.32     q12, r4          @ \n\t"
-      "vdup.32     q13, r4          @ \n\t"
-      "lsrs        r4, %2, #3       @ \n\t"
-      "beq         3f               @ \n\t"
-      "1:                           @ \n\t"
-      "vst1.32     {d24-d27}, [r0]! @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "bne         1b               @ \n\t"
-      "3:                           @ \n\t"
-      "ands        r4, %2, #7       @ \n\t"
-      "beq         5f               @ \n\t"
-      "4:                           @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "vst1.32     {d24[0]}, [r0]!  @ \n\t"
-      "bne         4b               @ \n\t"
-      "5:                           @ "
-      :
-      :"r" (x), "r"(caddr),"r"(n)
-      : "cc", "r0", "r4",  "memory",
-        "q12",
-        "d24", "d25", "d26", "d27"
-      );
+  long i = 0;
+
+  for(; i < n-4; i += 4)
+  {
+    x[i] = c;
+    x[i+1] = c;
+    x[i+2] = c;
+    x[i+3] = c;
+  }
+
+  for(; i < n; i++)
+    x[i] = c;
+
 }
 
 
 static void THFloatVector_diff_NEON(float *z, const float *x, const float *y, const ptrdiff_t n) {
-  __asm__ __volatile__ (
-      "mov         r0, %2           @ \n\t"
-      "mov         r1, %1           @ \n\t"
-      "mov         r2, %0           @ \n\t"
-      "lsrs        r4, %3, #3       @ \n\t"
-      "beq         3f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "1:                           @ \n\t"
-      "vsub.f32    q12, q8, q0      @ \n\t"
-      "vsub.f32    q13, q9, q1      @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "beq         2f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vst1.32     {d24-d27}, [r2]! @ \n\t"
-      "b           1b               @ \n\t"
-      "2:                           @ \n\t"
-      "vst1.32     {d24-d27}, [r2]! @ \n\t"
-      "3:                           @ \n\t"
-      "ands        r4, %3, #7       @ \n\t"
-      "beq         5f               @ \n\t"
-      "4:                           @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "vld1.32     {d16[0]}, [r1]!  @ \n\t"
-      "vld1.32     {d0[0]}, [r0]!   @ \n\t"
-      "vsub.f32    d24, d16, d0     @ \n\t"
-      "vst1.32     {d24[0]}, [r2]!  @ \n\t"
-      "bne         4b               @ \n\t"
-      "5:                           @ "
-      :
-      :"r" (z), "r" (x),"r" (y), "r"(n)
-      : "cc", "r0", "r1", "r2", "r4", "memory",
-        "q0", "q1", "q8", "q9", "q12", "q13",
-        "d0", "d1", "d2", "d3",
-        "d16", "d17", "d18", "d19", "d24", "d25", "d26", "d27"
-      );
+  long i = 0;
+
+  for(; i < n-4; i += 4)
+  {
+    z[i] = x[i] - y[i];
+    z[i+1] = x[i+1] - y[i+1];
+    z[i+2] = x[i+2] - y[i+2];
+    z[i+3] = x[i+3] - y[i+3];
+  }
+
+  for(; i < n; i++)
+    z[i] = x[i] - y[i];
+
 }
 
 
 static void THFloatVector_scale_NEON(float *y, const float c, const ptrdiff_t n) {
-  float ctemp = c;
-  float * caddr = &ctemp;
-  __asm__ __volatile__ (
-      "mov         r0, %0           @ \n\t"
-      "mov         r2, r0           @ \n\t"
-      "ldr         r5, [%1]         @ \n\t"
-      "vdup.32     q14, r5          @ \n\t"
-      "lsrs        r5, %2, #5       @ \n\t"
-      "beq         3f               @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vld1.32     {d4-d7}, [r0]!   @ \n\t"
-      "vld1.32     {d8-d11}, [r0]!  @ \n\t"
-      "vld1.32     {d12-d15}, [r0]! @ \n\t"
-      "1:                           @ \n\t"
-      "vmul.f32    q0, q0, q14      @ \n\t"
-      "vmul.f32    q1, q1, q14      @ \n\t"
-      "vmul.f32    q2, q2, q14      @ \n\t"
-      "vmul.f32    q3, q3, q14      @ \n\t"
-      "vmul.f32    q4, q4, q14      @ \n\t"
-      "vmul.f32    q5, q5, q14      @ \n\t"
-      "vmul.f32    q6, q6, q14      @ \n\t"
-      "vmul.f32    q7, q7, q14      @ \n\t"
-      "subs        r5, r5, #1       @ \n\t"
-      "beq         2f               @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vst1.32     {d4-d7}, [r2]!   @ \n\t"
-      "vld1.32     {d4-d7}, [r0]!   @ \n\t"
-      "vst1.32     {d8-d11}, [r2]!  @ \n\t"
-      "vld1.32     {d8-d11}, [r0]!  @ \n\t"
-      "vst1.32     {d12-d15}, [r2]! @ \n\t"
-      "vld1.32     {d12-d15}, [r0]! @ \n\t"
-      "b           1b               @ \n\t"
-      "2:                           @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "vst1.32     {d4-d7}, [r2]!   @ \n\t"
-      "vst1.32     {d8-d11}, [r2]!  @ \n\t"
-      "vst1.32     {d12-d15}, [r2]! @ \n\t"
-      "3:                           @ \n\t"
-      "lsrs        r5, %2, #4       @ \n\t"
-      "ands        r5, r5, #1       @ \n\t"
-      "beq         4f               @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vld1.32     {d4-d7}, [r0]!   @ \n\t"
-      "vmul.f32    q0, q0, q14      @ \n\t"
-      "vmul.f32    q1, q1, q14      @ \n\t"
-      "vmul.f32    q2, q2, q14      @ \n\t"
-      "vmul.f32    q3, q3, q14      @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "vst1.32     {d4-d7}, [r2]!   @ \n\t"
-      "4:                           @ \n\t"
-      "lsrs        r5, %2, #3       @ \n\t"
-      "ands        r5, r5, #1       @ \n\t"
-      "beq         5f               @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vmul.f32    q0, q0, q14      @ \n\t"
-      "vmul.f32    q1, q1, q14      @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "5:                           @ \n\t"
-      "ands        r5, %2, #7       @ \n\t"
-      "beq         7f               @ \n\t"
-      "6:                           @ \n\t"
-      "subs        r5, r5, #1       @ \n\t"
-      "vld1.32     d0[0], [r0]!     @ \n\t"
-      "vmul.f32    d0, d0, d28      @ \n\t"
-      "vst1.32     d0[0], [r2]!     @ \n\t"
-      "bne         6b               @ \n\t"
-      "7:                           @ "
-      :
-      :"r" (y), "r"(caddr),"r"(n)
-      : "cc", "r0", "r2", "r5", "memory",
-        "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q14",
-        "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
-        "d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15",
-        "d28", "d29"
-      );
+  long i = 0;
 
+  for(; i < n-4; i +=4)
+  {
+    y[i] *= c;
+    y[i+1] *= c;
+    y[i+2] *= c;
+    y[i+3] *= c;
+  }
+
+  for(; i < n; i++)
+    y[i] *= c;
 }
 
 static void THFloatVector_mul_NEON(float *y, const float *x, const ptrdiff_t n) {
-  __asm__ __volatile__ (
-      "mov         r0, %0           @ \n\t"
-      "mov         r1, %1           @ \n\t"
-      "mov         r2, r0           @ \n\t"
-      "lsrs        r4, %2, #3       @ \n\t"
-      "beq         3f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "1:                           @ \n\t"
-      "vmul.f32    q12, q8, q0      @ \n\t"
-      "vmul.f32    q13, q9, q1      @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "beq         2f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vst1.32     {d24-d27}, [r2]! @ \n\t"
-      "b           1b               @ \n\t"
-      "2:                           @ \n\t"
-      "vst1.32     {d24-d27}, [r2]! @ \n\t"
-      "3:                           @ \n\t"
-      "ands        r4, %2, #7       @ \n\t"
-      "beq         5f               @ \n\t"
-      "4:                           @ \n\t"
-      "subs        r4, r4, #1       @ \n\t"
-      "vld1.32     {d16[0]}, [r1]!  @ \n\t"
-      "vld1.32     {d0[0]}, [r0]!   @ \n\t"
-      "vmul.f32    q12, q8, q0      @ \n\t"
-      "vst1.32     {d24[0]}, [r2]!  @ \n\t"
-      "bne         4b               @ \n\t"
-      "5:                           @ "
-      :
-      :"r" (y),"r" (x),"r"(n)
-      : "cc", "r0", "r1", "r2", "r4", "memory",
-        "q0", "q1", "q8", "q9", "q12", "q13",
-        "d0", "d1", "d2", "d3",
-        "d16", "d17", "d18", "d19", "d24", "d25", "d26", "d27"
-      );
+  long i = 0;
+
+  for(; i < n-4; i += 4)
+  {
+    y[i] *= x[i];
+    y[i+1] *= x[i+1];
+    y[i+2] *= x[i+2];
+    y[i+3] *= x[i+3];
+  }
+
+  for(; i < n; i++)
+    y[i] *= x[i];
 }
 
 static void THFloatVector_add_NEON(float *y, const float *x, const float c, const ptrdiff_t n) {
-  float ctemp = c;
-  float * caddr = &ctemp;
-  __asm__ __volatile__ (
-      "mov         r0, %0           @ \n\t"
-      "mov         r1, %1           @ \n\t"
-      "mov         r2, r0           @ \n\t"
-      "ldr         r5, [%2]         @ \n\t"
-      "vdup.32     q14, r5          @ \n\t"
-      "lsrs        r5, %3, #4       @ \n\t"
-      "beq         3f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vld1.32     {d20-d23}, [r1]! @ \n\t"
-      "vld1.32     {d4-d7}, [r0]!   @ \n\t"
-      "1:                           @ \n\t"
-      "vmla.f32    q0, q8, q14      @ \n\t"
-      "vmla.f32    q1, q9, q14      @ \n\t"
-      "vmla.f32    q2, q10, q14     @ \n\t"
-      "vmla.f32    q3, q11, q14     @ \n\t"
-      "subs        r5, r5, #1       @ \n\t"
-      "beq         2f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d20-d23}, [r1]! @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vst1.32     {d4-d7}, [r2]!   @ \n\t"
-      "vld1.32     {d4-d7}, [r0]!   @ \n\t"
-      "b           1b               @ \n\t"
-      "2:                           @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "vst1.32     {d4-d7}, [r2]!   @ \n\t"
-      "3:                           @ \n\t"
-      "lsrs        r5, %3, #3       @ \n\t"
-      "ands        r5, #1           @ \n\t"
-      "beq         4f               @ \n\t"
-      "vld1.32     {d16-d19}, [r1]! @ \n\t"
-      "vld1.32     {d0-d3}, [r0]!   @ \n\t"
-      "vmla.f32    q0, q8, q14      @ \n\t"
-      "vmla.f32    q1, q9, q14      @ \n\t"
-      "vst1.32     {d0-d3}, [r2]!   @ \n\t"
-      "4:                           @ \n\t"
-      "ands        r5, %3, #7       @ \n\t"
-      "beq         6f               @ \n\t"
-      "5:                           @ \n\t"
-      "subs        r5, r5, #1       @ \n\t"
-      "vld1.32     {d16[0]}, [r1]!  @ \n\t"
-      "vld1.32     {d0[0]}, [r0]!   @ \n\t"
-      "vmla.f32    d0, d16, d28     @ \n\t"
-      "vst1.32     d0[0], [r2]!     @ \n\t"
-      "bne         5b               @ \n\t"
-      "6:                           @ "
-      :
-      :"r" (y),"r" (x), "r"(caddr),"r"(n)
-      : "cc", "r0", "r1", "r2", "r5", "memory",
-        "q0", "q1", "q2", "q3", "q14",
-        "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
-        "d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23", "d28", "d29"
-      );
+  long i = 0;
+
+  for(;i < n-4; i += 4)
+  {
+    y[i] += c * x[i];
+    y[i+1] += c * x[i+1];
+    y[i+2] += c * x[i+2];
+    y[i+3] += c * x[i+3];
+  }
+
+  for(; i < n; i++)
+    y[i] += c * x[i];
 }