commit aefaef35fabcecf8cbf2beda1d1a0673d2b8543d
author Marat Dukhan <maratek@google.com> Thu Apr 09 07:09:34 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Apr 09 07:12:09 2020 -0700
tree f01042c8079b43d29eca7a047256a5aff58cd121
parent 163a7e6fd68ec7d3be6b1ecdf44196214ced539a

Prepare xnn_params for variations in fused activations

- Define structures for a set of micro-kernels with the same fused activation
- Move existing micro-kernels into .minmax member

PiperOrigin-RevId: 305678097

src/convolution-nhwc.c

diff --git a/src/convolution-nhwc.c b/src/convolution-nhwc.c
index e0d73b3..f8c73ce 100644
--- a/src/convolution-nhwc.c
+++ b/src/convolution-nhwc.c
@@ -50,7 +50,7 @@
     size_t num_ukernels)
 {
   while (num_ukernels-- != 0) {
-    if (ukernel->mr == kernel_size) {
+    if (ukernel->primary_tile == kernel_size) {
       return ukernel;
     }
     ukernel++;
@@ -246,9 +246,9 @@
     case xnn_ukernel_type_dwconv:
     {
       assert(dwconv_parameters != NULL);
-      assert(dwconv_parameters->mr == kernel_size);
+      assert(dwconv_parameters->primary_tile == kernel_size);
 
-      const uint32_t c_stride = round_up_po2(groups, dwconv_parameters->cr);
+      const uint32_t c_stride = round_up_po2(groups, dwconv_parameters->channel_tile);
       const size_t packed_weights_size = (sizeof(uint8_t) * kernel_size + sizeof(int32_t)) * c_stride;
       convolution_op->packed_weights = xnn_allocate_simd_memory(packed_weights_size);
       if (convolution_op->packed_weights == NULL) {
@@ -259,21 +259,21 @@
       if (flags & XNN_FLAG_DEPTHWISE_CONVOLUTION) {
         xnn_pack_q8_dwconv_hwg_w(
           kernel_height, kernel_width,
-          groups, dwconv_parameters->cr,
+          groups, dwconv_parameters->channel_tile,
           input_zero_point, kernel_zero_point,
           kernel, bias, convolution_op->packed_weights);
       } else {
         xnn_pack_q8_dwconv_ghw_w(
           kernel_height, kernel_width,
-          groups, dwconv_parameters->cr,
+          groups, dwconv_parameters->channel_tile,
           input_zero_point, kernel_zero_point,
           kernel, bias, convolution_op->packed_weights);
       }
 
       convolution_op->ukernel.dwconv = (struct xnn_ukernel_dwconv) {
-        .unipass_function = dwconv_parameters->up,
-        .mr = dwconv_parameters->mr,
-        .qr = dwconv_parameters->qr,
+        .unipass_function = dwconv_parameters->minmax.unipass,
+        .primary_tile = dwconv_parameters->primary_tile,
+        .incremental_tile = dwconv_parameters->incremental_tile,
       };
 
       zero_size = sizeof(uint8_t) * c_stride + XNN_EXTRA_BYTES;
@@ -307,7 +307,7 @@
             .mr = xnn_params.q8.gemm.mr,
             .nr = nr,
             .kr = kr,
-            .general_case = xnn_params.q8.gemm.gemm,
+            .general_case = xnn_params.q8.gemm.minmax.gemm,
           };
           break;
         case xnn_ukernel_type_igemm:
@@ -328,7 +328,7 @@
             .mr = xnn_params.q8.gemm.mr,
             .nr = nr,
             .kr = kr,
-            .general_case = xnn_params.q8.gemm.igemm,
+            .general_case = xnn_params.q8.gemm.minmax.igemm,
           };
           break;
         default:
@@ -578,9 +578,9 @@
     case xnn_ukernel_type_dwconv:
     {
       assert(dwconv_parameters != NULL);
-      assert(dwconv_parameters->mr == kernel_size);
+      assert(dwconv_parameters->primary_tile == kernel_size);
 
-      const uint32_t c_stride = round_up_po2(groups, dwconv_parameters->cr);
+      const uint32_t c_stride = round_up_po2(groups, dwconv_parameters->channel_tile);
       const size_t packed_weights_size = (kernel_size + 1) * sizeof(float) * c_stride;
       convolution_op->packed_weights = xnn_allocate_simd_memory(packed_weights_size);
       if (convolution_op->packed_weights == NULL) {
@@ -591,19 +591,19 @@
       if (flags & XNN_FLAG_DEPTHWISE_CONVOLUTION) {
         xnn_pack_f32_dwconv_hwg_w(
           kernel_height, kernel_width,
-          groups, dwconv_parameters->cr,
+          groups, dwconv_parameters->channel_tile,
           kernel, bias, convolution_op->packed_weights);
       } else {
         xnn_pack_f32_dwconv_ghw_w(
           kernel_height, kernel_width,
-          groups, dwconv_parameters->cr,
+          groups, dwconv_parameters->channel_tile,
           kernel, bias, convolution_op->packed_weights);
       }
 
       convolution_op->ukernel.dwconv = (struct xnn_ukernel_dwconv) {
-        .unipass_function = dwconv_parameters->up,
-        .mr = dwconv_parameters->mr,
-        .qr = dwconv_parameters->qr,
+        .unipass_function = dwconv_parameters->minmax.unipass,
+        .primary_tile = dwconv_parameters->primary_tile,
+        .incremental_tile = dwconv_parameters->incremental_tile,
       };
 
       zero_size = sizeof(float) * c_stride;
@@ -636,8 +636,8 @@
             .mr = xnn_params.f32.gemm.mr,
             .nr = nr,
             .kr = kr,
-            .general_case = xnn_params.f32.gemm.gemm,
-            .mr1_case = xnn_params.f32.gemm.gemm1,
+            .general_case = xnn_params.f32.gemm.minmax.gemm,
+            .mr1_case = xnn_params.f32.gemm.minmax.gemm1,
           };
           break;
         case xnn_ukernel_type_igemm:
@@ -656,8 +656,8 @@
             .mr = xnn_params.f32.gemm.mr,
             .nr = nr,
             .kr = kr,
-            .general_case = xnn_params.f32.gemm.igemm,
-            .mr1_case = xnn_params.f32.gemm.igemm1,
+            .general_case = xnn_params.f32.gemm.minmax.igemm,
+            .mr1_case = xnn_params.f32.gemm.minmax.igemm1,
           };
           break;
         default: