From 0bcb0fcd981cb15329dfd4fb675097564164da18 Mon Sep 17 00:00:00 2001
From: tylov <tylo@norceresearch.no>
Date: Tue, 11 Jul 2023 22:16:09 +0200
Subject: Fixed an issue in template.h Reverted to cspan_md() and
 cspan_md_left() for column-major. Changed cspan_submdX(): add OutputSpanType
 as first parameter - aligns with cspan_slice() and adds type safety.

---
 include/stc/cspan.h         | 42 +++++++++++++++++++++++++-----------------
 include/stc/priv/template.h |  6 +++++-
 2 files changed, 30 insertions(+), 18 deletions(-)

(limited to 'include')

diff --git a/include/stc/cspan.h b/include/stc/cspan.h
index 027b5275..89986d6f 100644
--- a/include/stc/cspan.h
+++ b/include/stc/cspan.h
@@ -30,7 +30,7 @@ using_cspan(Intspan, int, 1);
 
 int demo1() {
     float raw[4*5];
-    Span2f ms = cspan_md('C', raw, 4, 5);
+    Span2f ms = cspan_md(raw, 4, 5);
     
     for (int i=0; i<ms.shape[0]; i++)
         for (int j=0; j<ms.shape[1]; j++)
@@ -147,26 +147,34 @@ using_cspan_tuple(7); using_cspan_tuple(8);
 
 
 // cspan_submd(): Reduce rank (N <= 4) Optimized, same as e.g. cspan_slice(Span2, &ms4, {x}, {y}, {c_ALL}, {c_ALL});
-#define cspan_submd2(self, x) \
-    {.data=cspan_at(self, x, 0), .shape={(self)->shape[1]}, .stride={.d={(self)->stride.d[1]}}}
-#define cspan_submd3(...) c_MACRO_OVERLOAD(cspan_submd3, __VA_ARGS__)
-#define cspan_submd3_2(self, x) \
-    {.data=cspan_at(self, x, 0, 0), .shape={(self)->shape[1], (self)->shape[2]}, \
+#define cspan_submd2(OutSpan, self, ...) _cspan_submdN(OutSpan, 2, self, __VA_ARGS__)
+#define cspan_submd3(OutSpan, self, ...) _cspan_submdN(OutSpan, 3, self, __VA_ARGS__)
+#define cspan_submd4(OutSpan, self, ...) _cspan_submdN(OutSpan, 4, self, __VA_ARGS__)
+
+#define _cspan_submdN(OutSpan, N, self, ...) \
+    _cspan_submd##N(c_static_assert(cspan_rank((OutSpan*)0) == N - c_NUMARGS(__VA_ARGS__)), self, __VA_ARGS__)
+
+#define _cspan_submd2(ok, self, x) \
+    {.data=cspan_at(self, x, 0) + ok, .shape={(self)->shape[1]}, .stride={.d={(self)->stride.d[1]}}}
+#define _cspan_submd3(...) c_MACRO_OVERLOAD(_cspan_submd3, __VA_ARGS__)
+#define _cspan_submd3_3(ok, self, x) \
+    {.data=cspan_at(self, x, 0, 0) + ok, .shape={(self)->shape[1], (self)->shape[2]}, \
                                     .stride={.d={(self)->stride.d[1], (self)->stride.d[2]}}}
-#define cspan_submd3_3(self, x, y) \
-    {.data=cspan_at(self, x, y, 0), .shape={(self)->shape[2]}, .stride={.d={(self)->stride.d[2]}}}
-#define cspan_submd4(...) c_MACRO_OVERLOAD(cspan_submd4, __VA_ARGS__)
-#define cspan_submd4_2(self, x) \
-    {.data=cspan_at(self, x, 0, 0, 0), .shape={(self)->shape[1], (self)->shape[2], (self)->shape[3]}, \
+#define _cspan_submd3_4(ok, self, x, y) \
+    {.data=cspan_at(self, x, y, 0) + ok, .shape={(self)->shape[2]}, .stride={.d={(self)->stride.d[2]}}}
+#define _cspan_submd4(...) c_MACRO_OVERLOAD(_cspan_submd4, __VA_ARGS__)
+#define _cspan_submd4_3(ok, self, x) \
+    {.data=cspan_at(self, x, 0, 0, 0) + ok, .shape={(self)->shape[1], (self)->shape[2], (self)->shape[3]}, \
                                        .stride={.d={(self)->stride.d[1], (self)->stride.d[2], (self)->stride.d[3]}}}
-#define cspan_submd4_3(self, x, y) \
-    {.data=cspan_at(self, x, y, 0, 0), .shape={(self)->shape[2], (self)->shape[3]}, \
+#define _cspan_submd4_4(ok, self, x, y) \
+    {.data=cspan_at(self, x, y, 0, 0) + ok, .shape={(self)->shape[2], (self)->shape[3]}, \
                                        .stride={.d={(self)->stride.d[2], (self)->stride.d[3]}}}
-#define cspan_submd4_4(self, x, y, z) \
-    {.data=cspan_at(self, x, y, z, 0), .shape={(self)->shape[3]}, .stride={.d={(self)->stride.d[3]}}}
+#define _cspan_submd4_5(ok, self, x, y, z) \
+    {.data=cspan_at(self, x, y, z, 0) + ok, .shape={(self)->shape[3]}, .stride={.d={(self)->stride.d[3]}}}
 
-
-#define cspan_md(order, array, ...) \
+#define cspan_md(array, ...) cspan_md_order('C', array, __VA_ARGS__)
+#define cspan_md_left(array, ...) cspan_md_order('F', array, __VA_ARGS__)
+#define cspan_md_order(order, array, ...) \
     {.data=array, .shape={__VA_ARGS__}, \
      .stride=*(c_PASTE(cspan_tuple, c_NUMARGS(__VA_ARGS__))*)_cspan_shape2stride(order, ((int32_t[]){__VA_ARGS__}), c_NUMARGS(__VA_ARGS__))}
 
diff --git a/include/stc/priv/template.h b/include/stc/priv/template.h
index 7138a87c..5551eeae 100644
--- a/include/stc/priv/template.h
+++ b/include/stc/priv/template.h
@@ -164,6 +164,8 @@
   #error "No i_key or i_val defined"
 #elif defined i_keyraw ^ defined i_keyto
   #error "Both i_keyraw/valraw and i_keyto/valto must be defined, if any"
+#elif defined i_keyfrom && !defined i_keyraw && !defined i_keyclone
+  #define i_keyclone i_keyfrom
 #elif defined i_keyfrom && !defined i_keyraw
   #error "i_keyfrom/valfrom defined without i_keyraw/valraw"
 #elif defined i_from || defined i_drop
@@ -261,7 +263,9 @@
   #error "i_val* must be defined for maps"
 #endif
 
-#if !(defined i_valclone || defined i_no_clone) && (defined i_valdrop || defined i_valraw)
+#if !defined i_valclone && defined i_valfrom && !defined i_valraw
+  #define i_valclone i_valfrom
+#elif !(defined i_valclone || defined i_no_clone) && (defined i_valdrop || defined i_valraw)
   #error i_valclone should be defined when i_valdrop or i_valraw is defined
 #endif
 #ifndef i_valraw
-- 
cgit v1.2.3