Remove unused parameter.

PiperOrigin-RevId: 766810743

Author: gonnet

src/microkernel-utils.c

@@ -14,7 +14,7 @@
 #include "src/xnnpack/math.h"
 
 static bool fits_in_cache(size_t mr, size_t nc, size_t m_stride,
-                          size_t n_stride, size_t cm_stride, size_t cn_stride,
+                          size_t n_stride, size_t cn_stride,
                           size_t cache_size, size_t cache_line_size) {
   // Check if the bytes fit.
   const size_t lines_mr = divide_round_up(mr * m_stride, cache_line_size);
@@ -32,7 +32,7 @@ static bool fits_in_cache(size_t mr, size_t nc, size_t m_stride,
 
 size_t xnn_gemm_best_tile_size(size_t num_groups, size_t m, size_t n,
                                size_t m_stride, size_t n_stride,
-                               size_t cm_stride, size_t cn_stride, size_t mr,
+                               size_t cn_stride, size_t mr,
                                size_t nr, size_t num_threads) {
   // Adjust `mr` and `nr` if they are larger than `m` and `n`, respectively.
   mr = min(mr, m);
@@ -55,11 +55,11 @@ size_t xnn_gemm_best_tile_size(size_t num_groups, size_t m, size_t n,
   size_t cache_size = hardware_config->l1_data_cache_bytes;
   size_t cache_line_size = hardware_config->l1_data_cache_line_size;
   if (XNN_ARCH_X86 || XNN_ARCH_X86_64 ||
-      (cache_size && !fits_in_cache(mr, nr, m_stride, n_stride, cm_stride,
+      (cache_size && !fits_in_cache(mr, nr, m_stride, n_stride, 
                                     cn_stride, cache_size, cache_line_size))) {
     cache_size = hardware_config->l2_data_cache_bytes;
     cache_line_size = hardware_config->l2_data_cache_line_size;
-    if (cache_size && !fits_in_cache(mr, nr, m_stride, n_stride, cm_stride,
+    if (cache_size && !fits_in_cache(mr, nr, m_stride, n_stride, 
                                      cn_stride, cache_size, cache_line_size)) {
       // Don't check for cache fit.
       cache_size = 0;
@@ -80,7 +80,7 @@ size_t xnn_gemm_best_tile_size(size_t num_groups, size_t m, size_t n,
     // If, however, we have more than one tile row, then we want the data used
     // to compute a tile of size `mr`x`j*nr` to fit in the cache.
     if (mr < m && cache_size &&
-        !fits_in_cache(mr, j * nr, m_stride, n_stride, cm_stride, cn_stride,
+        !fits_in_cache(mr, j * nr, m_stride, n_stride, cn_stride,
                        cache_size, cache_line_size)) {
       break;
     }

src/operators/batch-matrix-multiply-nc.c

@@ -908,9 +908,10 @@ static enum xnn_status reshape_batch_matrix_multiply_nc(
   // Compute the optimal tile size for this GEMM.
   const size_t nc = xnn_gemm_best_tile_size(
       /*num_groups=*/batch_size_c, m, n,
-      /*m_stride=*/batch_matrix_multiply_op->dynamic_context.gemm->gemm.a_stride,
-      /*n_stride=*/batch_matrix_multiply_op->dynamic_context.gemm->gemm.w_stride,
-      /*cm_stride=*/batch_matrix_multiply_op->dynamic_context.gemm->gemm.cm_stride,
+      /*m_stride=*/
+      batch_matrix_multiply_op->dynamic_context.gemm->gemm.a_stride,
+      /*n_stride=*/
+      batch_matrix_multiply_op->dynamic_context.gemm->gemm.w_stride,
       /*cn_stride=*/1 << log2_output_element_size, mr, nr, num_threads);
 
 #if XNN_MAX_UARCH_TYPES > 1

src/operators/convolution-nhwc.c

@@ -2044,7 +2044,6 @@ static enum xnn_status reshape_igemm(
           (input_width * convolution_op->input_pixel_stride
            << log2_input_element_size),
       /*n_stride=*/convolution_op->dynamic_context.igemm->igemm.w_stride,
-      /*cm_stride=*/convolution_op->dynamic_context.igemm->igemm.cm_stride,
       /*cn_stride=*/1 << log2_output_element_size, mr, nr, num_threads);
 
   if (dynamic_quantization && convolution_op->convolution_op->zero_size > 0) {

src/operators/deconvolution-nhwc.c

@@ -1340,7 +1340,6 @@ static enum xnn_status reshape_conv_path(
           (input_width * deconvolution_op->input_pixel_stride
            << log2_input_element_size),
       /*n_stride=*/deconvolution_op->dynamic_context.igemm->igemm.w_stride,
-      /*cm_stride=*/deconvolution_op->dynamic_context.igemm->igemm.cm_stride,
       /*cn_stride=*/1 << log2_output_element_size, mr, nr, num_threads);
 
   size_t igemm_compute_index = 0;

src/operators/dynamic-fully-connected-nc.c

@@ -551,10 +551,10 @@ static enum xnn_status reshape_dynamic_fully_connected_nc(
   // Compute the optimal tile size for this GEMM.
   const size_t nc = xnn_gemm_best_tile_size(
       /*num_groups=*/1, /*m=*/batch_size, /*n=*/output_channels,
-      /*m_stride=*/dynamic_fully_connected_op->dynamic_context.gemm->gemm.a_stride,
-      /*n_stride=*/dynamic_fully_connected_op->dynamic_context.gemm->gemm.w_stride,
-      /*cm_stride=*/
-      dynamic_fully_connected_op->dynamic_context.gemm->gemm.cm_stride,
+      /*m_stride=*/
+      dynamic_fully_connected_op->dynamic_context.gemm->gemm.a_stride,
+      /*n_stride=*/
+      dynamic_fully_connected_op->dynamic_context.gemm->gemm.w_stride,
       /*cn_stride=*/1 << log2_output_element_size, mr, nr,
       /*num_threads=*/pthreadpool_get_threads_count(threadpool));
 

src/operators/fully-connected-nc.c

@@ -2467,7 +2467,6 @@ static enum xnn_status reshape_fully_connected_nc(
       /*num_groups=*/1, /*m=*/batch_size, /*n=*/output_channels,
       /*m_stride=*/fully_connected_op->dynamic_context.gemm->gemm.a_stride,
       /*n_stride=*/fully_connected_op->dynamic_context.gemm->gemm.w_stride,
-      /*cm_stride=*/fully_connected_op->dynamic_context.gemm->gemm.cm_stride,
       /*cn_stride=*/1 << log2_output_element_size, mr, nr,
       /*num_threads=*/pthreadpool_get_threads_count(threadpool));
 

src/xnnpack/microkernel-utils.h

@@ -8,8 +8,6 @@
 
 #include <stddef.h>
 
-#include "src/xnnpack/common.h"
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -23,8 +21,8 @@ extern "C" {
 // fits in either the L1 or L2 cache.
 size_t xnn_gemm_best_tile_size(size_t num_groups, size_t m, size_t n,
                                size_t m_stride, size_t n_stride,
-                               size_t cm_stride, size_t cn_stride, size_t mr,
-                               size_t nr, size_t num_threads);
+                               size_t cn_stride, size_t mr, size_t nr,
+                               size_t num_threads);
 
 #ifdef __cplusplus
 }

test/microkernel-utils.cc

@@ -34,7 +34,7 @@ TEST(GEMM_BEST_TILE_SIZE, min_tiles_per_thread) {
          {(size_t)1, num_threads, 5 * num_threads, 10 * num_threads}) {
       const size_t nc = xnn_gemm_best_tile_size(
           num_groups, m, n, /*m_stride=*/k * sizeof(float),
-          /*n_stride=*/k * sizeof(float), /*cm_stride=*/n * sizeof(float),
+          /*n_stride=*/k * sizeof(float),
           /*cn_stride=*/sizeof(float), mr, nr, num_threads);
 
       // Check that `nc` is an integer multiple of `nr` if it is less than `n`.