Merge pull request #129 from foss-for-synopsys-dwc-arc-processors/mem_stride

[memory strides] add memory stride support to kernels

Author: JaccovG

build/rules.mk

@@ -119,6 +119,7 @@ else
 ifeq ($(TOOLCHAIN),mwdt)
 # place to add MWDT-specific common settings
 CFLAGS     +=-Hon=Long_enums
+CFLAGS     +=-mllvm -gen-lpcc=false
 DEPFLAGS    =-Hdepend=$(BUILD_DIR)/ -MD
 else
 $(error ERROR - Unsupported toolchain. Supported toolchains are 'gnu' and 'mwdt', default one is 'gnu')

include/mli_config.h

@@ -98,12 +98,18 @@
 */
 #if (ARC_PLATFORM == V2DSP_XY)
 #define MLI_PTR(p) __xy p *
+#define MLI_PTR_IS_XY true
 #define MLI_OUT_PTR(p) __xy p *
+#define MLI_OUT_PTR_IS_XY true
 #define MLI_CONV_OUT_PTR(p) p *
+#define MLI_CONV_OUT_PTR_IS_XY false
 #else
 #define MLI_PTR(p) p *
+#define MLI_PTR_IS_XY false
 #define MLI_OUT_PTR(p) p *
+#define MLI_OUT_PTR_IS_XY false
 #define MLI_CONV_OUT_PTR(p) p *
+#define MLI_CONV_OUT_PTR_IS_XY false
 #endif
 
 #endif // _MLI_CONFIG_H_

lib/gen/mli_krn_avepool_chw_func_body.txt

@@ -9,9 +9,13 @@
     int padding_bot = cfg->padding_bottom;
     int padding_left = cfg->padding_left;
     int padding_right = cfg->padding_right;
-    int channels_num = in->shape[FMAP_C_DIM_CHW];
     int kernel_height = cfg->kernel_height;
     int kernel_width = cfg->kernel_width;
+
+    // Define Data dimensions
+    auto in_prv = mli_prv_get_tensor_chw<MLI_PTR($d_type), MLI_PTR_IS_XY>(in,
+            $channels); // channels
+
     // assign hard coded values for this variation to some variables
 #if $stride_w
     MLI_CHECK_AND_FIX(stride_width, $stride_w);
@@ -31,20 +35,19 @@
 #if $kernel_h
     MLI_CHECK_AND_FIX(kernel_height, $kernel_h);
 #endif
-#if $channels
-    MLI_CHECK_AND_FIX(channels_num, $channels);
-#endif
-
-    // Data pointers
-    MLI_PTR($d_type) in_ftrs = (MLI_PTR($d_type ))in->data;
-    MLI_OUT_PTR($d_type) out_ftrs = (MLI_OUT_PTR($d_type ))out->data;
 
     // Define Data dimensions
-    const int in_height = in->shape[FMAP_H_DIM_CHW];
-    const int in_width = in->shape[FMAP_W_DIM_CHW];
+    const int out_width = CEIL_DIV(in_prv.width + padding_left + padding_right - kernel_width + 1, stride_width);
+    const int out_height = CEIL_DIV(in_prv.height + padding_top + padding_bot - kernel_height + 1, stride_height);
 
-    const int out_width = CEIL_DIV(in_width + padding_left + padding_right - kernel_width + 1, stride_width);
-    const int out_height = CEIL_DIV(in_height + padding_top + padding_bot - kernel_height + 1, stride_height);
+    // fill output tensor parameters
+    out->el_type = in->el_type;
+    out->rank = in->rank;
+    out->shape[FMAP_C_DIM_CHW] = in_prv.ch;
+    out->shape[FMAP_H_DIM_CHW] = out_height;
+    out->shape[FMAP_W_DIM_CHW] = out_width;
+    out->el_params.fx.frac_bits = in->el_params.fx.frac_bits;
+    const auto out_prv = mli_prv_get_tensor_chw<MLI_OUT_PTR($d_type), MLI_OUT_PTR_IS_XY>(out);
 
     const int row_beg = 0;
     const int row_end = out_height;
@@ -56,20 +59,10 @@
     $core_name(
         row_beg, row_end,
         clmn_beg, clmn_end,
-        in_ftrs, out_ftrs,
-        channels_num, in_width, in_height,
-        out_width, out_height,
+        in_prv, out_prv,
         kernel_height, kernel_width,
         stride_height, stride_width,
         padding_top, padding_left, padding_right, padding_bot);
 
-    // fill output tensor parameters
-    out->el_type = in->el_type;
-    out->rank = in->rank;
-    out->shape[FMAP_C_DIM_CHW] = channels_num;
-    out->shape[FMAP_H_DIM_CHW] = out_height;
-    out->shape[FMAP_W_DIM_CHW] = out_width;
-    out->el_params.fx.frac_bits = in->el_params.fx.frac_bits;
-
     return MLI_STATUS_OK;
 }

lib/gen/mli_krn_avepool_hwc_func_body.txt

@@ -9,9 +9,13 @@
     int padding_bot = cfg->padding_bottom;
     int padding_left = cfg->padding_left;
     int padding_right = cfg->padding_right;
-    int channels_num = in->shape[FMAP_C_DIM_HWC];
     int kernel_height = cfg->kernel_height;
     int kernel_width = cfg->kernel_width;
+
+    // Define Data dimensions
+    auto in_prv = mli_prv_get_tensor_hwc<MLI_PTR($d_type), MLI_PTR_IS_XY>(in,
+            $channels); // channels
+
     // assign hard coded values for this variation to some variables
 #if $stride_w
     MLI_CHECK_AND_FIX(stride_width, $stride_w);
@@ -31,20 +35,19 @@
 #if $kernel_h
     MLI_CHECK_AND_FIX(kernel_height, $kernel_h);
 #endif
-#if $channels
-    MLI_CHECK_AND_FIX(channels_num, $channels);
-#endif
-
-    // Data pointers
-    MLI_PTR($d_type) in_ftrs = (MLI_PTR($d_type ))in->data;
-    MLI_OUT_PTR($d_type) out_ftrs = (MLI_OUT_PTR($d_type ))out->data;
 
     // Define Data dimensions
-    const int in_height = in->shape[FMAP_H_DIM_HWC];
-    const int in_width = in->shape[FMAP_W_DIM_HWC];
+    const int out_width = CEIL_DIV(in_prv.width + padding_left + padding_right - kernel_width + 1, stride_width);
+    const int out_height = CEIL_DIV(in_prv.height + padding_top + padding_bot - kernel_height + 1, stride_height);
 
-    const int out_width = CEIL_DIV(in_width + padding_left + padding_right - kernel_width + 1, stride_width);
-    const int out_height = CEIL_DIV(in_height + padding_top + padding_bot - kernel_height + 1, stride_height);
+    // fill output tensor parameters
+    out->el_type = in->el_type;
+    out->rank = in->rank;
+    out->shape[FMAP_H_DIM_HWC] = out_height;
+    out->shape[FMAP_W_DIM_HWC] = out_width;
+    out->shape[FMAP_C_DIM_HWC] = in_prv.ch;
+    out->el_params.$el_params = in->el_params.$el_params;
+    const auto out_prv = mli_prv_get_tensor_hwc<MLI_OUT_PTR($d_type), MLI_OUT_PTR_IS_XY>(out);
 
     const int row_beg = 0;
     const int row_end = out_height;
@@ -56,20 +59,10 @@
     $core_name(
         row_beg, row_end,
         clmn_beg, clmn_end,
-        in_ftrs, out_ftrs,
-        channels_num, in_width, in_height,
-        out_width, out_height,
+        in_prv, out_prv,
         kernel_height, kernel_width,
         stride_height, stride_width,
         padding_top, padding_left, padding_right, padding_bot);
 
-    // fill output tensor parameters
-    out->el_type = in->el_type;
-    out->rank = in->rank;
-    out->shape[FMAP_H_DIM_HWC] = out_height;
-    out->shape[FMAP_W_DIM_HWC] = out_width;
-    out->shape[FMAP_C_DIM_HWC] = channels_num;
-    out->el_params.$el_params = in->el_params.$el_params;
-
     return MLI_STATUS_OK;
 }

lib/gen/mli_krn_conv2d_nhwc_func_body.txt

@@ -11,10 +11,15 @@
     int padding_bot = cfg->padding_bottom;
     int padding_left = cfg->padding_left;
     int padding_right = cfg->padding_right;
-    int kernel_height = weights->shape[KRNL_H_DIM_HWC];
-    int kernel_width = weights->shape[KRNL_W_DIM_HWC];
-    int out_ch = weights->shape[KRNL_C_DIM_HWC];
-    int in_ch = in->shape[FMAP_C_DIM_HWC];
+
+    // Define Data dimensions
+    const auto in_prv = mli_prv_get_tensor_hwc<MLI_PTR($d_type), MLI_PTR_IS_XY>(in,
+            $channels); // channels
+    const auto w = mli_prv_get_conv2d_weights_tensor_nhwc<MLI_PTR($w_type), MLI_PTR_IS_XY>(weights,
+            $channels,  // channels
+            $kernel_w,  // kernel_width
+            $kernel_h); // kernel_height
+    __builtin_assume(in_prv.ch == w.in_ch);
 
     // assign hard coded values for this variation to some variables
 #if $stride_w
@@ -29,15 +34,6 @@
     MLI_CHECK_AND_FIX(padding_left, $padding_left);
     MLI_CHECK_AND_FIX(padding_right, $padding_right);
 #endif
-#if $kernel_w
-    MLI_CHECK_AND_FIX(kernel_width, $kernel_w);
-#endif
-#if $kernel_h
-    MLI_CHECK_AND_FIX(kernel_height, $kernel_h);
-#endif
-#if $channels
-    MLI_CHECK_AND_FIX(in_ch, $channels);
-#endif
 
     mli_minmax_t val_limit;
     // fill output tensor el_type parameter
@@ -46,17 +42,18 @@
     val_limit = mli_prv_get_relu_min_max(&cfg->relu, out);
 
     // Data pointers
-    MLI_PTR($d_type) in_ftrs = (MLI_PTR($d_type ))in->data;
-    MLI_CONV_OUT_PTR($d_type) out_ftrs = (MLI_CONV_OUT_PTR($d_type ))out->data;
-    MLI_PTR($w_type) wt = (MLI_PTR($w_type ))weights->data;
     MLI_PTR($b_type) bs = (MLI_PTR($b_type ))bias->data;
 
     // Define Data dimensions
-    int in_height = in->shape[FMAP_H_DIM_HWC];
-    int in_width = in->shape[FMAP_W_DIM_HWC];
+    int out_width = CEIL_DIV(in_prv.width + padding_left + padding_right - w.kernel_width + 1, stride_width);
+    int out_height = CEIL_DIV(in_prv.height + padding_top + padding_bot - w.kernel_height + 1, stride_height);
 
-    int out_width = CEIL_DIV(in_width + padding_left + padding_right - kernel_width + 1, stride_width);
-    int out_height = CEIL_DIV(in_height + padding_top + padding_bot - kernel_height + 1, stride_height);
+    // fill output tensor parameters
+    out->rank = in->rank;
+    out->shape[FMAP_C_DIM_HWC] = w.out_ch;
+    out->shape[FMAP_H_DIM_HWC] = out_height;
+    out->shape[FMAP_W_DIM_HWC] = out_width;
+    const auto out_prv = mli_prv_get_tensor_hwc<MLI_CONV_OUT_PTR($d_type), MLI_CONV_OUT_PTR_IS_XY>(out);
 
     // Define quantization specific params
     s8asym_quant_specific_params params;
@@ -70,16 +67,11 @@
     cent_area.clmn_end = out_width;
 
     $core_name<$d_type, $w_type, $b_type, int32_t>(
-            in_ftrs, wt, bs, out_ftrs, &cent_area, params,
-            (int8_t)val_limit.min, (int8_t)val_limit.max, in_ch, in_width, in_height, 
-            out_ch, out_width, out_height, kernel_height, kernel_width,
-            stride_height, stride_width, padding_top, padding_left, padding_bot, padding_right);
-
-    // fill output tensor parameters
-    out->rank = in->rank;
-    out->shape[FMAP_C_DIM_HWC] = out_ch;
-    out->shape[FMAP_H_DIM_HWC] = out_height;
-    out->shape[FMAP_W_DIM_HWC] = out_width;
+        in_prv, w, bs, out_prv,
+        &cent_area, params,
+        (int8_t)val_limit.min, (int8_t)val_limit.max,
+        stride_height, stride_width,
+        padding_top, padding_left, padding_bot, padding_right);
 
     return MLI_STATUS_OK;
 }

lib/gen/mli_krn_depthwise_conv2d_hwcn_func_body.txt

@@ -11,10 +11,14 @@
     int padding_bot = cfg->padding_bottom;
     int padding_left = cfg->padding_left;
     int padding_right = cfg->padding_right;
-    int kernel_height = weights->shape[KRNL_DW_H_DIM_HWC];
-    int kernel_width = weights->shape[KRNL_DW_W_DIM_HWC];
-    int out_ch = weights->shape[KRNL_DW_C_DIM_HWC];
-    int in_ch = in->shape[FMAP_C_DIM_HWC];
+
+    // Define Data dimensions
+    const auto in_prv = mli_prv_get_tensor_hwc<MLI_PTR($d_type), MLI_PTR_IS_XY>(in,
+            $channels); // channels
+    const auto w = mli_prv_get_conv2d_weights_tensor_1hwn<MLI_PTR($w_type), MLI_PTR_IS_XY>(
+            weights,
+            $kernel_w,  // kernel_width
+            $kernel_h); // kernel_height
 
     // assign hard coded values for this variation to some variables
 #if $stride_w
@@ -29,15 +33,6 @@
     MLI_CHECK_AND_FIX(padding_left, $padding_left);
     MLI_CHECK_AND_FIX(padding_right, $padding_right);
 #endif
-#if $kernel_w
-    MLI_CHECK_AND_FIX(kernel_width, $kernel_w);
-#endif
-#if $kernel_h
-    MLI_CHECK_AND_FIX(kernel_height, $kernel_h);
-#endif
-#if $channels
-    MLI_CHECK_AND_FIX(in_ch, $channels);
-#endif
 
     mli_minmax_t val_limit;
     // fill output tensor el_type parameter
@@ -46,18 +41,18 @@
     val_limit = mli_prv_get_relu_min_max(&cfg->relu, out);
 
     // Data pointers
-    MLI_PTR($d_type) in_ftrs = (MLI_PTR($d_type ))in->data;
-    MLI_CONV_OUT_PTR($d_type) out_ftrs = (MLI_CONV_OUT_PTR($d_type ))out->data;
-    MLI_PTR($w_type) wt = (MLI_PTR($w_type ))weights->data;
     MLI_PTR($b_type) bs = (MLI_PTR($b_type ))bias->data;
 
      // Define Data dimensions
-    int in_height = in->shape[FMAP_H_DIM_HWC];
-    int in_width = in->shape[FMAP_W_DIM_HWC];
+    int out_width = CEIL_DIV(in_prv.width + padding_left + padding_right - w.kernel_width + 1, stride_width);
+    int out_height = CEIL_DIV(in_prv.height + padding_top + padding_bot - w.kernel_height + 1, stride_height);
 
-
-    int out_width = CEIL_DIV(in_width + padding_left + padding_right - kernel_width + 1, stride_width);
-    int out_height = CEIL_DIV(in_height + padding_top + padding_bot - kernel_height + 1, stride_height);
+    // fill output tensor parameters
+    out->rank = in->rank;
+    out->shape[FMAP_C_DIM_HWC] = w.out_ch;
+    out->shape[FMAP_H_DIM_HWC] = out_height;
+    out->shape[FMAP_W_DIM_HWC] = out_width;
+    const auto out_prv = mli_prv_get_tensor_hwc<MLI_CONV_OUT_PTR($d_type), MLI_CONV_OUT_PTR_IS_XY>(out);
 
     // Define quantization specific params
     s8asym_quant_specific_params params;
@@ -71,16 +66,9 @@
     cent_area.clmn_end = out_width;
 
     $core_name<$d_type, $w_type, $b_type, int32_t>(
-            in_ftrs, wt, bs, out_ftrs, &cent_area, params,
-            (int8_t)val_limit.min, (int8_t)val_limit.max, in_ch, in_width, in_height, 
-            out_ch, out_width, out_height, kernel_height, kernel_width,
+            in_prv, w, bs, out_prv, &cent_area, params,
+            (int8_t)val_limit.min, (int8_t)val_limit.max,
             stride_height, stride_width, padding_top, padding_left, padding_bot, padding_right);
 
-    // fill output tensor parameters
-    out->rank = in->rank;
-    out->shape[FMAP_C_DIM_HWC] = out_ch;
-    out->shape[FMAP_H_DIM_HWC] = out_height;
-    out->shape[FMAP_W_DIM_HWC] = out_width;
-
     return MLI_STATUS_OK;
 }

lib/gen/mli_krn_maxpool_chw_func_body.txt

@@ -9,9 +9,13 @@
     int32_t padding_bot = cfg->padding_bottom;
     int32_t padding_left = cfg->padding_left;
     int32_t padding_right = cfg->padding_right;
-    int32_t channels_num = in->shape[FMAP_C_DIM_CHW];
     int32_t kernel_height = cfg->kernel_height;
     int32_t kernel_width = cfg->kernel_width;
+
+    // Define Data dimensions
+    auto in_prv = mli_prv_get_tensor_chw<MLI_PTR($d_type), MLI_PTR_IS_XY>(in,
+            $channels); // channels
+
     // assign hard coded values for this variation to some variables
 #if $stride_w
     MLI_CHECK_AND_FIX(stride_width, $stride_w);
@@ -31,20 +35,19 @@
 #if $kernel_h
     MLI_CHECK_AND_FIX(kernel_height, $kernel_h);
 #endif
-#if $channels
-    MLI_CHECK_AND_FIX(channels_num, $channels);
-#endif
-
-    // Data pointers
-    MLI_PTR($d_type) in_ftrs = (MLI_PTR($d_type ))in->data;
-    MLI_OUT_PTR($d_type) out_ftrs = (MLI_OUT_PTR($d_type ))out->data;
 
     // Define Data dimensions
-    const int32_t in_height = in->shape[FMAP_H_DIM_CHW];
-    const int32_t in_width = in->shape[FMAP_W_DIM_CHW];
+    const int32_t out_width = CEIL_DIV(in_prv.width + padding_left + padding_right - kernel_width + 1, stride_width);
+    const int32_t out_height = CEIL_DIV(in_prv.height + padding_top + padding_bot - kernel_height + 1, stride_height);
 
-    const int32_t out_width = CEIL_DIV(in_width + padding_left + padding_right - kernel_width + 1, stride_width);
-    const int32_t out_height = CEIL_DIV(in_height + padding_top + padding_bot - kernel_height + 1, stride_height);
+    // fill output tensor parameters
+    out->el_type = in->el_type;
+    out->rank = in->rank;
+    out->shape[FMAP_C_DIM_CHW] = in_prv.ch;
+    out->shape[FMAP_H_DIM_CHW] = out_height;
+    out->shape[FMAP_W_DIM_CHW] = out_width;
+    out->el_params.fx.frac_bits = in->el_params.fx.frac_bits;
+    const auto out_prv = mli_prv_get_tensor_chw<MLI_OUT_PTR($d_type), MLI_OUT_PTR_IS_XY>(out);
 
     const int32_t row_beg = 0;
     const int32_t row_end = out_height;
@@ -54,23 +57,13 @@
     mli_prv_fx_init_dsp_ctrl();
 
     $core_name(
-        in_ftrs, out_ftrs,
+        in_prv, out_prv,
         row_beg, row_end,
         clmn_beg, clmn_end,
-        channels_num, in_width, in_height,
-        out_width, out_height,
         kernel_height, kernel_width,
         stride_height, stride_width,
         padding_top, padding_bot, padding_left, padding_right,
         $kernelpadding);
 
-    // fill output tensor parameters
-    out->el_type = in->el_type;
-    out->rank = in->rank;
-    out->shape[FMAP_C_DIM_CHW] = channels_num;
-    out->shape[FMAP_H_DIM_CHW] = out_height;
-    out->shape[FMAP_W_DIM_CHW] = out_width;
-    out->el_params.fx.frac_bits = in->el_params.fx.frac_bits;
-
     return MLI_STATUS_OK;
 }