Add SFPU binary bitwise, int32 add, div

tenstorrent · Dec 4, 2024 · ed52a73 · ed52a73
1 parent a7620d0
commit ed52a73
Show file tree

Hide file tree

Showing 4 changed files with 62 additions and 4 deletions.
diff --git a/common/inc/ckernel_sfpu.h b/common/inc/ckernel_sfpu.h
@@ -14,6 +14,7 @@
 #include "sfpu/ckernel_sfpu_abs.h"
 #include "sfpu/ckernel_sfpu_add_int32.h"
 #include "sfpu/ckernel_sfpu_binary.h"
+#include "sfpu/ckernel_sfpu_binary_bitwise.h"
 #include "sfpu/ckernel_sfpu_cast_fp32_to_fp16a.h"
 #include "sfpu/ckernel_sfpu_clamp.h"
 #include "sfpu/ckernel_sfpu_comp.h"

diff --git a/common/inc/sfpu/ckernel_sfpu_add_int32.h b/common/inc/sfpu/ckernel_sfpu_add_int32.h
@@ -17,7 +17,7 @@ namespace sfpu
 {
 
 template <bool APPROXIMATION_MODE, int ITERATIONS>
-inline void _add_int32_(const int iterations, const uint dst_offset) {
+inline void _add_int32_(const uint dst_offset) {
     // Operand A is input1 (int32)
     // Operand B is input2 (int32)
     // Output is int32

diff --git a/common/inc/sfpu/ckernel_sfpu_binary.h b/common/inc/sfpu/ckernel_sfpu_binary.h
@@ -111,10 +111,17 @@ inline void _calculate_sfpu_binary_(const uint dst_offset)
             result = in0 * in1;
         } else if constexpr (BINOP_MODE == DIV_BINARY) {
             v_if (in1 == 0) {
-                result = std::numeric_limits<float>::infinity();
-                result = setsgn(result, in0);
+                v_if (in0 == 0) {
+                    result = std::numeric_limits<float>::quiet_NaN();
+                } v_else {
+                    result = std::numeric_limits<float>::infinity();
+                    result = setsgn(result, in0);
+                }
+                v_endif;
+            } v_elseif (in0 == in1) {
+                result = vConst1;
             } v_else {
-                result = in0 * setsgn(_sfpu_reciprocal_(in1), in1);
+                result = in0 * setsgn(_sfpu_reciprocal_<4>(in1), in1);
             }
             v_endif;
         } else if constexpr (BINOP_MODE == RSUB_BINARY) {

diff --git a/common/inc/sfpu/ckernel_sfpu_binary_bitwise.h b/common/inc/sfpu/ckernel_sfpu_binary_bitwise.h
@@ -0,0 +1,50 @@
+// SPDX-FileCopyrightText: © 2024 Tenstorrent AI ULC
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "ckernel.h"
+#include "ckernel_defs.h"
+#include "noc_nonblocking_api.h"
+#include "sfpi.h"
+#include <limits.h>
+
+using namespace sfpi;
+
+namespace ckernel
+{
+namespace sfpu
+{
+
+enum {
+    AND_BINARY = 0,
+    OR_BINARY = 1,
+    XOR_BINARY = 2,
+};  // BITWISE_MODE
+
+template <bool APPROXIMATION_MODE, int BITWISE_MODE, int ITERATIONS = 8>
+inline void _calculate_sfpu_binary_bitwise_(const uint dst_offset)
+{
+    // SFPU microcode
+    for (int d = 0; d < ITERATIONS; d++) {
+        constexpr uint dst_tile_size = 64;
+
+        TTI_SFPLOAD(0,12,ADDR_MOD_7,0);
+        TT_SFPLOAD(1,12,ADDR_MOD_7,dst_offset*dst_tile_size);
+
+        if constexpr (BITWISE_MODE == AND_BINARY) {
+            TTI_SFPAND(0,1,0,0);
+        } else if constexpr (BITWISE_MODE == OR_BINARY) {
+            TTI_SFPOR(0,1,0,0);
+        } else if constexpr (BITWISE_MODE == XOR_BINARY) {
+            TTI_SFPXOR(0,1,0,0);
+        }
+
+        TTI_SFPSTORE(0,12,ADDR_MOD_7,0);
+        dst_reg++;
+    }
+}
+
+}  // namespace sfpu
+}  // namespace ckernel