Add binary sfpu div init (#50)

* Add binary sfpu div init * Add enum classes for op type template param
tenstorrent · Jan 2, 2025 · a2645b5 · a2645b5
1 parent 33a7f6a
commit a2645b5
Show file tree

Hide file tree

Showing 2 changed files with 32 additions and 24 deletions.
diff --git a/common/inc/sfpu/ckernel_sfpu_binary.h b/common/inc/sfpu/ckernel_sfpu_binary.h
@@ -17,14 +17,14 @@ namespace ckernel
 namespace sfpu
 {
 
-enum {
-    ADD_BINARY = 0,
-    SUB_BINARY = 1,
-    MUL_BINARY = 2,
-    DIV_BINARY = 3,
-    RSUB_BINARY = 4,
-    POW_BINARY = 5
-};  // BINOP_MODE
+enum class BinaryOp : uint8_t {
+    ADD = 0,
+    SUB = 1,
+    MUL = 2,
+    DIV = 3,
+    RSUB = 4,
+    POW = 5
+};
 
 sfpi_inline vFloat _calculate_sfpu_binary_power_(vFloat base, vFloat pow)
 {
@@ -93,7 +93,7 @@ sfpi_inline vFloat _calculate_sfpu_binary_power_(vFloat base, vFloat pow)
     return result;
 }
 
-template <bool APPROXIMATION_MODE, int BINOP_MODE, int ITERATIONS = 8>
+template <bool APPROXIMATION_MODE, BinaryOp BINOP, int ITERATIONS = 8>
 inline void _calculate_sfpu_binary_(const uint dst_offset)
 {
     // SFPU microcode
@@ -103,13 +103,13 @@ inline void _calculate_sfpu_binary_(const uint dst_offset)
         vFloat in1 = dst_reg[dst_offset * dst_tile_size];
         vFloat result = 0.0f;
 
-        if constexpr (BINOP_MODE == ADD_BINARY) {
+        if constexpr (BINOP == BinaryOp::ADD) {
             result = in0 + in1;
-        } else if constexpr (BINOP_MODE == SUB_BINARY) {
+        } else if constexpr (BINOP == BinaryOp::SUB) {
             result = in0 - in1;
-        } else if constexpr (BINOP_MODE == MUL_BINARY) {
+        } else if constexpr (BINOP == BinaryOp::MUL) {
             result = in0 * in1;
-        } else if constexpr (BINOP_MODE == DIV_BINARY) {
+        } else if constexpr (BINOP == BinaryOp::DIV) {
             v_if (in1 == 0) {
                 v_if (in0 == 0) {
                     result = std::numeric_limits<float>::quiet_NaN();
@@ -124,9 +124,9 @@ inline void _calculate_sfpu_binary_(const uint dst_offset)
                 result = in0 * setsgn(_sfpu_reciprocal_<4>(in1), in1);
             }
             v_endif;
-        } else if constexpr (BINOP_MODE == RSUB_BINARY) {
+        } else if constexpr (BINOP == BinaryOp::RSUB) {
             result = in1 - in0;
-        } else if constexpr (BINOP_MODE == POW_BINARY) {
+        } else if constexpr (BINOP == BinaryOp::POW) {
             result = _calculate_sfpu_binary_power_(in0, in1);
         }
 
@@ -135,5 +135,13 @@ inline void _calculate_sfpu_binary_(const uint dst_offset)
     }
 }
 
+template <bool APPROXIMATION_MODE /*unused*/, BinaryOp BINOP>
+inline void _sfpu_binary_init_()
+{
+    if constexpr (BINOP == BinaryOp::DIV) {
+        _init_reciprocal_<APPROXIMATION_MODE>();
+    }
+}
+
 }  // namespace sfpu
 }  // namespace ckernel
diff --git a/common/inc/sfpu/ckernel_sfpu_binary_bitwise.h b/common/inc/sfpu/ckernel_sfpu_binary_bitwise.h
@@ -17,13 +17,13 @@ namespace ckernel
 namespace sfpu
 {
 
-enum {
-    AND_BINARY = 0,
-    OR_BINARY = 1,
-    XOR_BINARY = 2,
-};  // BITWISE_MODE
+enum class BinaryBitwiseOp : uint8_t {
+    AND = 0,
+    OR = 1,
+    XOR = 2,
+};
 
-template <bool APPROXIMATION_MODE, int BITWISE_MODE, int ITERATIONS = 8>
+template <bool APPROXIMATION_MODE, BinaryBitwiseOp BITWISE_OP, int ITERATIONS = 8>
 inline void _calculate_sfpu_binary_bitwise_(const uint dst_offset)
 {
     // SFPU microcode
@@ -33,11 +33,11 @@ inline void _calculate_sfpu_binary_bitwise_(const uint dst_offset)
         TTI_SFPLOAD(0,4,3,0);
         TT_SFPLOAD(1,4,3,dst_offset*dst_tile_size);
 
-        if constexpr (BITWISE_MODE == AND_BINARY) {
+        if constexpr (BITWISE_OP == BinaryBitwiseOp::AND) {
             TTI_SFPAND(0,1,0,0);
-        } else if constexpr (BITWISE_MODE == OR_BINARY) {
+        } else if constexpr (BITWISE_OP == BinaryBitwiseOp::OR) {
             TTI_SFPOR(0,1,0,0);
-        } else if constexpr (BITWISE_MODE == XOR_BINARY) {
+        } else if constexpr (BITWISE_OP == BinaryBitwiseOp::XOR) {
             TTI_SFPXOR(0,1,0,0);
         }