#16174: Support for int32 subtraction for WHB0 and BH (#17359)

### Ticket
#16174 

### Problem description
Subtraction of int32 dtype was not supported on WHB0 and BH.

### What's changed
Added necessary APIs, LLKs, and modified codepaths to include sub int32
operation.
This operation was done through SFPU. 

### Checklist
- [x] Post commit CI passes
- [x] Blackhole Post commit (if applicable)
- [ ] Model regression CI testing passes (if applicable)
- [ ] Device performance regression CI testing passes (if applicable)
- [ ] **(For models and ops writers)** Full [new
models](https://github.com/tenstorrent/tt-metal/actions/workflows/full-new-models-suite.yaml)
tests passes
- [x] New/Existing tests provide coverage for changes

tests/ttnn/unit_tests/operations/eltwise/test_binary_fp32.py

@@ -93,6 +93,26 @@ def test_add_int32(device, ttnn_function):
     assert status
 
 
+@skip_for_grayskull("Unsupported dtype for Grayskull")
+@pytest.mark.parametrize(
+    "ttnn_function",
+    [
+        ttnn.sub,
+    ],
+)
+def test_sub_int32(device, ttnn_function):
+    x_torch = torch.tensor([[11, 23, 0, -23, -1, -100]], dtype=torch.int32)
+    y_torch = torch.tensor([[78, 99, 34, -33, -1, 100]], dtype=torch.int32)
+    golden_fn = ttnn.get_golden_function(ttnn_function)
+    z_torch = golden_fn(x_torch, y_torch)
+    x_tt = ttnn.from_torch(x_torch, dtype=ttnn.int32, layout=ttnn.TILE_LAYOUT, device=device)
+    y_tt = ttnn.from_torch(y_torch, dtype=ttnn.int32, layout=ttnn.TILE_LAYOUT, device=device)
+    z_tt = ttnn.from_torch(z_torch, dtype=ttnn.int32, layout=ttnn.TILE_LAYOUT, device=device)
+    z_tt_sub = ttnn.sub(x_tt, y_tt)
+    tt_out = ttnn.to_torch(z_tt_sub)
+    assert torch.allclose(z_torch, tt_out, atol=1e-10, rtol=1e-5, equal_nan=False)
+
+
 @skip_for_grayskull("Unsupported dtype for Grayskull")
 @pytest.mark.parametrize(
     "ttnn_function",

tt_metal/hw/ckernels/blackhole/metal/llk_api/llk_sfpu/ckernel_sfpu_sub_int32.h

@@ -0,0 +1,22 @@
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "ckernel.h"
+#include "ckernel_defs.h"
+#include "sfpi.h"
+
+using namespace sfpi;
+
+namespace ckernel {
+namespace sfpu {
+
+template <bool APPROXIMATION_MODE, bool SIGN_MAGNITUDE_FORMAT, int ITERATIONS = 8>
+inline void calculate_sub_int32(const uint dst_offset) {
+    _sub_int32_<APPROXIMATION_MODE, SIGN_MAGNITUDE_FORMAT, ITERATIONS>(dst_offset);
+}
+
+}  // namespace sfpu
+}  // namespace ckernel

tt_metal/hw/ckernels/blackhole/metal/llk_api/llk_sfpu/llk_math_eltwise_binary_sfpu_sub_int32.h

@@ -0,0 +1,27 @@
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "llk_math_eltwise_binary_sfpu_init.h"
+#include "llk_math_eltwise_binary_sfpu_params.h"
+#include "ckernel_sfpu_sub_int32.h"
+
+namespace ckernel {
+
+// New LLK SFPU APIs
+
+template <bool APPROXIMATE>
+inline void llk_math_eltwise_binary_sfpu_sub_int32_init() {
+    llk_math_eltwise_binary_sfpu_init<SfpuType::unused, APPROXIMATE>();
+}
+
+template <bool APPROXIMATE, bool SIGN_MAGNITUDE_FORMAT>
+inline void llk_math_eltwise_binary_sfpu_sub_int32(
+    uint dst_index0, uint32_t dst_index1, int vector_mode = VectorMode::RC) {
+    llk_math_eltwise_binary_sfpu_params<APPROXIMATE>(
+        ckernel::sfpu::calculate_sub_int32<APPROXIMATE, SIGN_MAGNITUDE_FORMAT>, dst_index0, dst_index1, vector_mode);
+}
+
+}  // namespace ckernel

tt_metal/hw/ckernels/wormhole_b0/metal/llk_api/llk_sfpu/ckernel_sfpu_sub_int32.h

@@ -0,0 +1,22 @@
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "ckernel.h"
+#include "ckernel_defs.h"
+#include "sfpi.h"
+
+using namespace sfpi;
+
+namespace ckernel {
+namespace sfpu {
+
+template <bool APPROXIMATION_MODE, bool SIGN_MAGNITUDE_FORMAT, int ITERATIONS = 8>
+inline void calculate_sub_int32(const uint dst_offset) {
+    _sub_int32_<APPROXIMATION_MODE, SIGN_MAGNITUDE_FORMAT, ITERATIONS>(dst_offset);
+}
+
+}  // namespace sfpu
+}  // namespace ckernel

tt_metal/hw/ckernels/wormhole_b0/metal/llk_api/llk_sfpu/llk_math_eltwise_binary_sfpu_sub_int32.h

@@ -0,0 +1,27 @@
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "llk_math_eltwise_binary_sfpu_init.h"
+#include "llk_math_eltwise_binary_sfpu_params.h"
+#include "ckernel_sfpu_sub_int32.h"
+
+namespace ckernel {
+
+// New LLK SFPU APIs
+
+template <bool APPROXIMATE>
+inline void llk_math_eltwise_binary_sfpu_sub_int32_init() {
+    llk_math_eltwise_binary_sfpu_init<SfpuType::unused, APPROXIMATE>();
+}
+
+template <bool APPROXIMATE, bool SIGN_MAGNITUDE_FORMAT>
+inline void llk_math_eltwise_binary_sfpu_sub_int32(
+    uint dst_index0, uint32_t dst_index1, int vector_mode = VectorMode::RC) {
+    llk_math_eltwise_binary_sfpu_params<APPROXIMATE>(
+        ckernel::sfpu::calculate_sub_int32<APPROXIMATE, SIGN_MAGNITUDE_FORMAT>, dst_index0, dst_index1, vector_mode);
+}
+
+}  // namespace ckernel

tt_metal/include/compute_kernel_api/sub_int32_sfpu.h

@@ -0,0 +1,47 @@
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "compute_kernel_api/common_globals.h"
+#ifdef TRISC_MATH
+#include "llk_math_eltwise_binary_sfpu_sub_int32.h"
+#define MAIN math_main()
+#define MATH(x) x
+#else
+#define MATH(x)
+#endif
+
+namespace ckernel {
+
+// clang-format off
+/**
+ * Performs an elementwise sub operation with the two integer inputs: y = sub(x0,x1)
+ * Output overwrites first operand in DST.
+ *
+ * The DST register buffer must be in acquired state via *acquire_dst* call. This call is blocking and is only available
+ * on the compute engine.
+ * A maximum of 4 tiles from each operand can be loaded into DST at once, for a total of 8 tiles,
+ * when using 16 bit formats. This gets reduced to 2 tiles from each operand for 32 bit formats.
+ *
+ * Return value: None
+ *
+ * | Argument              | Description                                                                 | Type     | Valid Range                                           | Required |
+ * |-----------------------|-----------------------------------------------------------------------------|----------|-------------------------------------------------------|----------|
+ * | idst0                 | The index of the tile in DST register buffer to use as first operand        | uint32_t | Must be less than the size of the DST register buffer | True     |
+ * | idst1                 | The index of the tile in DST register buffer to use as second operand       | uint32_t | Must be less than the size of the DST register buffer | True     |
+ * | sign_magnitude_format | Whether the Int32 values are in sign-magnitude format (not 2's complement)  | bool     |                                                       | False    |
+ */
+// clang-format on
+template <bool sign_magnitude_format = false>
+ALWI void sub_int32_tile(uint32_t idst0, uint32_t idst1) {
+    MATH((llk_math_eltwise_binary_sfpu_sub_int32<APPROX, sign_magnitude_format>(idst0, idst1)));
+}
+
+/**
+ * Please refer to documentation for any_init.
+ */
+ALWI void sub_int32_tile_init() { MATH((llk_math_eltwise_binary_sfpu_sub_int32_init<APPROX>())); }
+
+}  // namespace ckernel

ttnn/cpp/ttnn/operations/eltwise/binary/common/binary_op_utils.cpp

@@ -191,8 +191,13 @@ std::map<std::string, std::string> get_defines_fp32(
             }
             break;
         case BinaryOpType::SUB:
-            new_defines.insert({"BINOP_INIT", fmt::format("sub_binary_tile_init();")});
-            op_name = "sub_binary_tile";
+            if (input_a_dtype == DataType::INT32 && input_b_dtype == DataType::INT32) {
+                new_defines.insert({"SUB_INT32_INIT", "sub_int32_tile_init();"});
+                op_name = "sub_int32_tile";
+            } else {
+                new_defines.insert({"BINOP_INIT", "sub_binary_tile_init();"});
+                op_name = "sub_binary_tile";
+            }
             break;
         case BinaryOpType::MUL:
             new_defines.insert({"BINOP_INIT", fmt::format("mul_binary_tile_init();")});

ttnn/cpp/ttnn/operations/eltwise/binary/device/binary_device_operation.cpp

@@ -17,8 +17,9 @@ namespace ttnn::operations::binary {
 namespace utils {
     bool is_binary_sfpu_op(BinaryOpType val, DataType a, DataType b) {
     switch (val) {
-        case BinaryOpType::ADD: return ((a == DataType::FLOAT32 && b == DataType::FLOAT32) || (a == DataType::INT32 && b == DataType::INT32));
+        case BinaryOpType::ADD:
         case BinaryOpType::SUB:
+            return ((a == DataType::FLOAT32 && b == DataType::FLOAT32) || (a == DataType::INT32 && b == DataType::INT32));
         case BinaryOpType::MUL:
         case BinaryOpType::DIV_FAST:
         case BinaryOpType::RSUB:

ttnn/cpp/ttnn/operations/eltwise/binary/device/kernels/compute/eltwise_binary_sfpu_kernel.cpp

@@ -13,6 +13,7 @@
 #include "compute_kernel_api/binary_bitwise_sfpu.h"
 #include "compute_kernel_api/binary_shift.h"
 #include "compute_kernel_api/add_int32_sfpu.h"
+#include "compute_kernel_api/sub_int32_sfpu.h"
 
 #define PRE_SCALE defined SFPU_OP_INIT_PRE_IN0_0 || defined SFPU_OP_INIT_PRE_IN1_0
 
@@ -113,6 +114,9 @@ void MAIN {
 #ifdef ADD_INT32_INIT
             ADD_INT32_INIT
 #endif
+#ifdef SUB_INT32_INIT
+            SUB_INT32_INIT
+#endif
 #ifdef BITWISE_INIT
             BITWISE_INIT
 #endif