fix: unary_zero_16m

Author: RivinHD

CMakeLists.txt

@@ -110,8 +110,8 @@ set(KERNEL_FILES
     br_matmul_lt16_lt4nRest_k.cpp
 
     unary/unary_all.h
-    unary/unary_zero_4m_n.h
-    unary/unary_zero_4m_n.cpp
+    unary/unary_zero_16m_n.h
+    unary/unary_zero_16m_n.cpp
 )
 
 set(ARM_INSTRUCTION_FILES
@@ -174,7 +174,7 @@ set(TEST_KERNELS
 
     unary/unary.test.h
     unary/unary.test.cpp
-    unary/unary_zero_4m_n.test.cpp
+    unary/unary_zero_16m_n.test.cpp
 )
 
 set(TEST_ARM_INSTRUCTION_FILES

src/main/kernels/unary/unary_all.h

@@ -1,6 +1,6 @@
 #ifndef MINI_JIT_KERNELS_UNARY_ALL_H
 #define MINI_JIT_KERNELS_UNARY_ALL_H
 
-#include "unary_zero_4m_n.h"
+#include "unary_zero_16m_n.h"
 
 #endif  // MINI_JIT_KERNELS_UNARY_ALL_H

…c/main/kernels/unary/unary_zero_4m_n.cpp …/main/kernels/unary/unary_zero_16m_n.cppsrc/main/kernels/unary/unary_zero_4m_n.cpp renamed to src/main/kernels/unary/unary_zero_16m_n.cpp src/main/kernels/unary/unary_zero_4m_n.cpp renamed to src/main/kernels/unary/unary_zero_16m_n.cpp

@@ -1,12 +1,12 @@
-#include "unary_zero_4m_n.h"
+#include "unary_zero_16m_n.h"
 #include "../../arm_instructions/arm_all.h"
 
-void mini_jit::kernels::unary_zero_4m_n(mini_jit::Kernel &kernel, const uint32_t m_loop_4, const uint32_t n_loop,
-                                        const bool use_init_and_end)
+void mini_jit::kernels::unary_zero_16m_n(mini_jit::Kernel &kernel, const uint32_t m_loop_16, const uint32_t n_loop,
+                                         const bool use_init_and_end)
 {
   using namespace mini_jit::arm_instructions;
 
-  release_assert(m_loop_4 != 0, "Cannot use a matrix with a m loop of size zero.");
+  release_assert(m_loop_16 != 0, "Cannot use a matrix with a m loop of size zero.");
   release_assert(n_loop != 0, "Cannot use a matrix with a n loop of size zero.");
 
   if (use_init_and_end)
@@ -40,8 +40,8 @@ void mini_jit::kernels::unary_zero_4m_n(mini_jit::Kernel &kernel, const uint32_t
       // Offset the used leading dimension by the size of floats
       lsl(x3, x3, 2),  // x3 * 4 = x3 * sizeof(float)
 
-      mov(x8, x1),      // Store the inital value of x1, to be restored in the N loop
-      mov(x9, 4 * 16),  // 4 * 16Byte Hold the number of bytes that are stored in the loop
+      mov(x8, x1),         // Store the inital value of x1, to be restored in the N loop
+      mov(x9, 4 * 4 * 4),  // 4 * 4 * sizeof(float) Hold the number of bytes that are stored in the loop
 
       // Zero four register so we can fill the matrix with zeros
       eor(v0, t16b, v0, t16b, v0, t16b),  // Zero the v0 register
@@ -57,10 +57,10 @@ void mini_jit::kernels::unary_zero_4m_n(mini_jit::Kernel &kernel, const uint32_t
     // loop over n
     sub(x16, x16, 1),
 
-    mov(x1, x9),  // Restore x1 for the m loop
+    mov(x1, x8),  // Restore x1 for the m loop
 
     // x17 iterator for the m_loop
-    mov(x17, m_loop_4),
+    mov(x17, m_loop_16),
     // loop over m
     sub(x17, x17, 1),
 
@@ -70,7 +70,7 @@ void mini_jit::kernels::unary_zero_4m_n(mini_jit::Kernel &kernel, const uint32_t
     cbnz(x17, -2 * 4),
 
     // Updates for the matrix B
-    add(x9, x3, x9),  // lda + initial position
+    add(x8, x3, x8),  // lda + initial position
 
     // loop back to n
     cbnz(x16, -7 * 4),

src/main/kernels/unary/unary_zero_4m_n.h src/main/kernels/unary/unary_zero_16m_n.hsrc/main/kernels/unary/unary_zero_4m_n.h renamed to src/main/kernels/unary/unary_zero_16m_n.h

@@ -1,5 +1,5 @@
-#ifndef MINI_JIT_KERNELS_UNARY_ZERO_M_4N_H
-#define MINI_JIT_KERNELS_UNARY_ZERO_M_4N_H
+#ifndef MINI_JIT_KERNELS_UNARY_ZERO_M_16N_H
+#define MINI_JIT_KERNELS_UNARY_ZERO_M_16N_H
 
 #include "../../Kernel.h"
 #include <cstdint>
@@ -12,11 +12,11 @@ namespace mini_jit
      * @brief Generates a M x 4*N unary zero kernel.
      *
      * @param kernel The kernel to add instructions too.
-     * @param m_loop_4 The repetitions of the m block of size.
+     * @param m_loop_16 The repetitions of the m block of size.
      * @param n_loop The repetitions of the n block of size 4.
      * @param use_init_and_end Indicates if the procedural call standard, initializing setup and the ret instruction are used. Defaults to
      */
-    void unary_zero_4m_n(mini_jit::Kernel &kernel, const uint32_t m_loop_4, const uint32_t n_loop, const bool use_init_and_end = true);
+    void unary_zero_16m_n(mini_jit::Kernel &kernel, const uint32_t m_loop_16, const uint32_t n_loop, const bool use_init_and_end = true);
 
   }  // namespace kernels
 }  // namespace mini_jit

src/main/kernels/unary/unary_zero_m_4n.bin

@@ -1,17 +1,17 @@
-#include "../../../main/kernels/unary/unary_zero_4m_n.h"
+#include "../../../main/kernels/unary/unary_zero_16m_n.h"
 #include "unary.test.h"
 #include <catch2/catch_test_macros.hpp>
 #include <catch2/generators/catch_generators.hpp>
 #include <catch2/generators/catch_generators_range.hpp>
 #include <cstdint>
 
-TEST_CASE("Test unary zero_4m_n jited correctness random data", "[jit][correctness][gemm]")
+TEST_CASE("Test unary zero_16m_n jited correctness random data", "[jit][correctness][gemm]")
 {
-  const uint32_t M = GENERATE(16, 64, 512, 2048);
-  const uint32_t N = GENERATE(16, 50, 64, 512, 2048);
+  const uint32_t M = GENERATE(64, 512, 2048);
+  const uint32_t N = GENERATE(50, 64, 512, 2048);
   CAPTURE(M, N);
   UnaryTestFixture unaryTest(M, N);
   unaryTest.SetUp(TestInfill::Random);
-  mini_jit::kernels::unary_zero_4m_n(unaryTest.native_kernel, M / 4, N);
+  mini_jit::kernels::unary_zero_16m_n(unaryTest.native_kernel, M / 16, N);
   unaryTest.RunTest(M, M, UnaryType::Zero);
 }