test loads

clamchowder · Oct 8, 2023 · 0f93a07 · 0f93a07
1 parent 3c53bd6
commit 0f93a07
Show file tree

Hide file tree

Showing 2 changed files with 56 additions and 0 deletions.
diff --git a/InstructionRate/x86_instructionrate.c b/InstructionRate/x86_instructionrate.c
@@ -84,6 +84,7 @@ extern uint64_t mul16(uint64_t iterations) __attribute((sysv_abi));
 extern uint64_t mul64(uint64_t iterations) __attribute((sysv_abi));
 extern uint64_t load128(uint64_t iterations, int *arr) __attribute((sysv_abi));
 extern uint64_t spacedload128(uint64_t iterations, int *arr) __attribute((sysv_abi));
+extern uint64_t loadscalar(uint64_t iterations, int *arr) __attribute((sysv_abi));
 extern uint64_t load256(uint64_t iterations, float *arr) __attribute((sysv_abi));
 extern uint64_t load512(uint64_t iterations, float *arr) __attribute((sysv_abi));
 extern uint64_t store128(uint64_t iterations, int *arr, int *sink) __attribute((sysv_abi));
@@ -125,6 +126,7 @@ int *intTestArr;
 int intSinkArr[8] __attribute__ ((aligned (64))) = { 2, 3, 4, 5, 6, 7, 8, 9 };
 
 uint64_t load128wrapper(uint64_t iterations) __attribute((sysv_abi));
+uint64_t loadscalarwrapper(uint64_t iterations) __attribute((sysv_abi));
 uint64_t spacedload128wrapper(uint64_t iterations) __attribute((sysv_abi));
 uint64_t spacedstorescalarwrapper(uint64_t iterations) __attribute((sysv_abi));
 uint64_t load256wrapper(uint64_t iterations) __attribute((sysv_abi));
@@ -420,6 +422,8 @@ int main(int argc, char *argv[]) {
     printf("2:1 mixed 16-bit/64-bit imul per clk: %.2f\n", measureFunction(iterations, clockSpeedGhz, mixmul16mul64_21));
 
   // load/store
+  if (argc == 1 || argc > 1 && strncmp(argv[1], "loadscalar", 10) == 0)
+    printf("64-bit scalar loads per clk: %.2f\n", measureFunction(iterations, clockSpeedGhz, loadscalarwrapper));
   if (argc == 1 || argc > 1 && strncmp(argv[1], "load128", 7) == 0)
     printf("128-bit loads per clk: %.2f\n", measureFunction(iterations, clockSpeedGhz, load128wrapper));
   if (argc == 1 || argc > 1 && strncmp(argv[1], "spacedload128", 13) == 0)
@@ -471,6 +475,10 @@ __attribute((sysv_abi)) uint64_t load256wrapper(uint64_t iterations) {
   return load256(iterations, fpTestArr);
 }
 
+__attribute((sysv_abi)) uint64_t loadscalarwrapper(uint64_t iterations) {
+  return loadscalar(iterations, intTestArr);
+}
+
 __attribute((sysv_abi)) uint64_t load512wrapper(uint64_t iterations) {
   return load512(iterations, fpTestArr);
 }

diff --git a/InstructionRate/x86_instructionrate.s b/InstructionRate/x86_instructionrate.s
@@ -84,6 +84,7 @@
 .global store128
 .global store256
 .global store512
+.global loadscalar
 .global spacedstorescalar
 .global mixaddmul128int
 .global mixmul16mul64
@@ -3984,6 +3985,53 @@ mixmul16mul64_21_loop:
   pop %rbx
   ret
 
+loadscalar:
+  push %rbx
+  push %rcx
+  push %r8
+  push %r9
+  push %r10
+  push %r11
+  push %r12
+  push %r13
+  push %r14
+  push %r15
+  mov $20, %r9
+loadscalar_loop:
+  mov (%rsi), %r15
+  mov 8(%rsi), %r14
+  mov 16(%rsi), %r13
+  mov 24(%rsi), %r12
+  mov 32(%rsi), %r11
+  mov 40(%rsi), %r10
+  mov 48(%rsi), %r15
+  mov 56(%rsi), %r14
+  mov 64(%rsi), %r13
+  mov 72(%rsi), %r12 
+  mov 80(%rsi), %r11
+  mov 88(%rsi), %r10
+  mov 96(%rsi), %r15
+  mov 104(%rsi), %r14
+  mov 112(%rsi), %r13  
+  mov 120(%rsi), %r12
+  mov 128(%rsi), %r11
+  mov 136(%rsi), %r10
+  mov 144(%rsi), %r15
+  mov 152(%rsi), %r14   
+  sub %r9, %rdi
+  jnz loadscalar_loop
+  pop %r15
+  pop %r14
+  pop %r13
+  pop %r12
+  pop %r11
+  pop %r10
+  pop %r9
+  pop %r8
+  pop %rcx
+  pop %rbx
+  ret 
+
 spacedstorescalar:
   push %rbx
   push %rcx