Add prelu support for blackhole

tt_metal/hw/ckernels/blackhole/metal/llk_api/llk_math_unary_sfpu_api.h

@@ -25,3 +25,4 @@
 #include "llk_math_eltwise_unary_sfpu_trigonometry.h"
 #include "llk_math_eltwise_unary_sfpu_unary_comp.h"
 #include "llk_math_eltwise_unary_sfpu_fill.h"
+#include "llk_math_eltwise_unary_sfpu_prelu.h"

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

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "ckernel.h"
+#include "ckernel_defs.h"
+#include "noc_nonblocking_api.h"
+#include "ckernel_sfpu_converter.h"
+
+
+using namespace sfpi;
+
+namespace ckernel {
+namespace sfpu {
+
+template <bool APPROXIMATION_MODE, int ITERATIONS = 8>
+inline void calculate_fill(const uint value) {
+
+    // SFPU microcode
+    Converter c_value;
+    c_value.u = value;
+    vFloat fill_val = c_value.f;
+
+    #pragma GCC unroll 0
+    for (int d = 0; d < ITERATIONS; d++)
+    {
+        vFloat a = dst_reg[0];
+        v_if(a < 0.0f) {
+            a = a * init;
+        }
+        v_endif;
+        dst_reg[0] = a;
+        dst_reg++;
+    }
+}
+}  // namespace sfpu
+}  // namespace ckernel

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

@@ -0,0 +1,29 @@
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "ckernel_sfpu_prelu.h"
+#include "llk_math_eltwise_unary_sfpu_params.h"
+#include "llk_math_eltwise_unary_sfpu_init.h"
+
+namespace ckernel {
+
+// New LLK SFPU APIs
+
+template <bool APPROXIMATE>
+inline void llk_math_eltwise_unary_sfpu_prelu_init() {
+    llk_math_eltwise_unary_sfpu_init<SfpuType::prelu, APPROXIMATE>();
+}
+
+template <bool APPROXIMATE>
+inline void llk_math_eltwise_unary_sfpu_prelu(uint dst_index, uint param0, int vector_mode = (int)VectorMode::RC) {
+    llk_math_eltwise_unary_sfpu_params<APPROXIMATE>(
+        ckernel::sfpu::calculate_prelu<APPROXIMATE>,
+        dst_index,
+        vector_mode,
+        param0);
+}
+
+}  // namespace ckernel

tt_metal/hw/ckernels/blackhole/metal/llk_api/llk_sfpu_types.h

@@ -87,5 +87,6 @@ enum SfpuType {
     ceil,
     unused,
     cumsum,
-    fill
+    fill,
+    prelu,
 };

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

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
 //
 // SPDX-License-Identifier: Apache-2.0
 
@@ -23,7 +23,8 @@ inline void calculate_prelu(uint value) {
     c_value.u = value;
     vFloat init = c_value.f;
 
-    for (int d = 0; d < 8; d++)
+    #pragma GCC unroll 8
+    for (int d = 0; d < ITERATIONS; d++)
     {
         vFloat a = dst_reg[0];
         v_if(a < 0.0f) {