From c51dc6098c2e3789c4b97b4737b8b79f9178602c Mon Sep 17 00:00:00 2001
From: Quaternions <krakow20@gmail.com>
Date: Wed, 4 Sep 2024 12:03:12 -0700
Subject: [PATCH] matrix: directly implement dot product to avoid a copy

---
 fixed_wide_vectors/src/macros/fixed_wide.rs | 35 ++++++++++-----------
 1 file changed, 16 insertions(+), 19 deletions(-)

diff --git a/fixed_wide_vectors/src/macros/fixed_wide.rs b/fixed_wide_vectors/src/macros/fixed_wide.rs
index 96e43c1..da74531 100644
--- a/fixed_wide_vectors/src/macros/fixed_wide.rs
+++ b/fixed_wide_vectors/src/macros/fixed_wide.rs
@@ -87,16 +87,14 @@ macro_rules! impl_wide_vector_operations {
 #[macro_export(local_inner_macros)]
 macro_rules! impl_matrix_wide_mul_transpose_helper {
 	(
-		$value:ident,
+		$lhs_axis:expr, $wide_mul:ident, $rhs:ident,
 		($struct: ident { $($field: ident), + }),
 		($from_struct: ident { $($from_field: ident), + }),
 		$static_field: ident
 	) => {
-		$struct {
-			$(
-				$field: $value.$from_field.$static_field
-			), +
-		}
+		$crate::sum_repeating!(
+			$( + $lhs_axis.$field.$wide_mul($rhs.$from_field.$static_field) ) +
+		)
 	}
 }
 #[doc(hidden)]
@@ -104,7 +102,7 @@ macro_rules! impl_matrix_wide_mul_transpose_helper {
 macro_rules! impl_matrix_wide_mul_inner {
 	(
 		// MatY<VecX>.MatX<VecZ> = MatY<VecZ>
-		$lhs:ident, $lhs_field_outer:ident, $wide_dot:ident, $rhs:ident,
+		$lhs:ident, $lhs_field_outer:ident, $wide_mul:ident, $rhs:ident,
 		$struct_inner_thru: tt, //VecX
 		($struct_inner: ident { $($field_inner: ident), + }), //VecX
 		($rhs_struct_inner: ident { $($rhs_field_inner: ident), + }), //VecZ
@@ -112,15 +110,14 @@ macro_rules! impl_matrix_wide_mul_inner {
 	) => {
 		$rhs_struct_inner {
 			$(
-				$rhs_field_inner: $lhs.$lhs_field_outer.$wide_dot(
-					//construct a transposed vector with the same width as $struct_outer
-					$crate::impl_matrix_wide_mul_transpose_helper!{
-						$rhs,
-						$struct_inner_thru, //VecZ
-						$rhs_outer, //MatX
-						$rhs_field_inner
-					}
-				)
+				//directly dot product to avoid a copy
+				$rhs_field_inner: $crate::impl_matrix_wide_mul_transpose_helper!{
+					//lhs_axis
+					$lhs.$lhs_field_outer,$wide_mul,$rhs,
+					$struct_inner_thru, //VecZ
+					$rhs_outer, //MatX
+					$rhs_field_inner
+				}
 			), +
 		}
 	}
@@ -130,7 +127,7 @@ macro_rules! impl_matrix_wide_mul_inner {
 macro_rules! impl_matrix_wide_mul_outer {
 	(
 		// MatY<VecX>.MatX<VecZ> = MatY<VecZ>
-		$lhs:ident, $wide_dot:ident, $rhs:ident,
+		$lhs:ident, $wide_mul:ident, $rhs:ident,
 		//result matrix shape
 		($struct_outer: ident { $($field_outer: ident), + }),//MatY
 		$rhs_struct_inner: tt,//VecZ
@@ -142,7 +139,7 @@ macro_rules! impl_matrix_wide_mul_outer {
 		$struct_outer {
 			$(
 				$field_outer: $crate::impl_matrix_wide_mul_inner!{
-					$lhs, $field_outer, $wide_dot, $rhs,
+					$lhs, $field_outer, $wide_mul, $rhs,
 					$struct_inner, //VecX
 					$struct_inner, //VecX
 					$rhs_struct_inner, //VecZ
@@ -182,7 +179,7 @@ macro_rules! impl_matrix_wide_mul {
 			pub fn [<wide_dot_ $size_outer x $size_inner _ $size_inner x $rhs_size_inner _ $lhs _ $rhs>](self,rhs:$matrix_inner<$rhs_struct_inner<fixed_wide::fixed::Fixed<{$rhs},{$rhs*32}>>>)->$struct_outer<$rhs_struct_inner<fixed_wide::fixed::Fixed<{$lhs+$rhs},{($lhs+$rhs)*32}>>>{
 				$crate::impl_matrix_wide_mul_outer!(
 					//constituent idents
-					self,[<wide_dot_ $lhs _ $rhs>],rhs,
+					self,[<wide_mul_ $lhs _ $rhs>],rhs,
 					//result matrix shape
 					($struct_outer { $($field_outer), + }),
 					($rhs_struct_inner { $($rhs_field_inner), + }),