skcms→019d877 add a basic profiling mode to src/Transform.c

Change-Id: I136064f8b772f1aa5ecfedac4b0f9ebb92da9d73
Reviewed-on: https://skia-review.googlesource.com/121486
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>

third_party/skcms/src/Transform.c

@@ -7,6 +7,7 @@
 
 #include "../skcms.h"
 #include "LinearAlgebra.h"
+#include "Macros.h"
 #include "PortableMath.h"
 #include "TransferFunction.h"
 #include "Transform.h"
@@ -15,6 +16,47 @@
 #include <stdint.h>
 #include <string.h>
 
+#if defined(SKCMS_PROFILE)
+    // Basic profiling tools to time each Op.  Not at all thread safe.
+
+    #include <stdio.h>
+    #include <stdlib.h>
+
+    #define M(op) +1
+    static uint64_t cycles[FOREACH_Op(M)];
+    #undef M
+
+    static void profile_dump_stats() {
+    #define M(op) #op,
+        static const char* names[] = { FOREACH_Op(M) };
+    #undef M
+        for (int i = 0; i < ARRAY_COUNT(cycles); i++) {
+            if (cycles[i]) {
+                fprintf(stderr, "%16s: %12llu cycles\n", names[i], cycles[i]);
+            }
+        }
+    }
+
+    static Op profile_next_op(Op op) {
+        static uint64_t start    = 0;
+        static uint64_t* current = NULL;
+
+        if (!current) {
+            atexit(profile_dump_stats);
+        } else {
+            *current += __builtin_readcyclecounter() - start;
+        }
+
+        current = &cycles[op];
+        start   = __builtin_readcyclecounter();
+        return op;
+    }
+#else
+    static inline Op profile_next_op(Op op) {
+        return op;
+    }
+#endif
+
 #if defined(__clang__)
     typedef float    __attribute__((ext_vector_type(4)))   Fx4;
     typedef int32_t  __attribute__((ext_vector_type(4))) I32x4;

third_party/skcms/src/Transform.h

@@ -9,57 +9,56 @@
 
 #pragma once
 
+#define FOREACH_Op(M) \
+    M(noop)           \
+    M(load_565)       \
+    M(load_888)       \
+    M(load_8888)      \
+    M(load_1010102)   \
+    M(load_161616)    \
+    M(load_16161616)  \
+    M(load_hhh)       \
+    M(load_hhhh)      \
+    M(load_fff)       \
+    M(load_ffff)      \
+    M(swap_rb)        \
+    M(clamp)          \
+    M(invert)         \
+    M(force_opaque)   \
+    M(premul)         \
+    M(unpremul)       \
+    M(matrix_3x3)     \
+    M(matrix_3x4)     \
+    M(lab_to_xyz)     \
+    M(tf_r)           \
+    M(tf_g)           \
+    M(tf_b)           \
+    M(tf_a)           \
+    M(table_8_r)      \
+    M(table_8_g)      \
+    M(table_8_b)      \
+    M(table_8_a)      \
+    M(table_16_r)     \
+    M(table_16_g)     \
+    M(table_16_b)     \
+    M(table_16_a)     \
+    M(clut_3D_8)      \
+    M(clut_3D_16)     \
+    M(clut_4D_8)      \
+    M(clut_4D_16)     \
+    M(store_565)      \
+    M(store_888)      \
+    M(store_8888)     \
+    M(store_1010102)  \
+    M(store_161616)   \
+    M(store_16161616) \
+    M(store_hhh)      \
+    M(store_hhhh)     \
+    M(store_fff)      \
+    M(store_ffff)
+
 typedef enum {
-    Op_noop,
-
-    Op_load_565,
-    Op_load_888,
-    Op_load_8888,
-    Op_load_1010102,
-    Op_load_161616,
-    Op_load_16161616,
-    Op_load_hhh,
-    Op_load_hhhh,
-    Op_load_fff,
-    Op_load_ffff,
-
-    Op_swap_rb,
-    Op_clamp,
-    Op_invert,
-    Op_force_opaque,
-    Op_premul,
-    Op_unpremul,
-
-    Op_matrix_3x3,
-    Op_matrix_3x4,
-    Op_lab_to_xyz,
-
-    Op_tf_r,
-    Op_tf_g,
-    Op_tf_b,
-    Op_tf_a,
-    Op_table_8_r,
-    Op_table_8_g,
-    Op_table_8_b,
-    Op_table_8_a,
-    Op_table_16_r,
-    Op_table_16_g,
-    Op_table_16_b,
-    Op_table_16_a,
-
-    Op_clut_3D_8,
-    Op_clut_3D_16,
-    Op_clut_4D_8,
-    Op_clut_4D_16,
-
-    Op_store_565,
-    Op_store_888,
-    Op_store_8888,
-    Op_store_1010102,
-    Op_store_161616,
-    Op_store_16161616,
-    Op_store_hhh,
-    Op_store_hhhh,
-    Op_store_fff,
-    Op_store_ffff,
+    #define M(op) Op_##op,
+    FOREACH_Op(M)
+    #undef M
 } Op;

third_party/skcms/src/Transform_inl.h

@@ -570,7 +570,7 @@ static void NS(exec_ops)(const Op* ops, const void** args,
                          const char* src, char* dst, int i) {
     F r = F0, g = F0, b = F0, a = F0;
     while (true) {
-        switch (*ops++) {
+        switch (profile_next_op(*ops++)) {
             case Op_noop: break;
 
             case Op_load_565:{