Improving 4bit quant mat mul performance by shifting position of -8 operation.

Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime.

Differential Revision: D85721578

Author: trivedivivek

backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_tiled.glsl

@@ -86,6 +86,13 @@ void main() {
         VEC4_T mat1_vec4 = VEC4_T(texelFetch(t_in, ivec3(txpos, out_row + i, 0), 0));
       $for j in range(4):
           mat1[i * 4 + ${j}] = mat1_vec4[${j}];
+
+      $if QUANT_NBITS == 4:
+        // Apply -8 * mat1 bias here rather then below, to effectively reduce overall number of math operations performed during runtime.
+        const T accum = mat1[i * 4 + 0] + mat1[i * 4 + 1] + mat1[i * 4 + 2] + mat1[i * 4 + 3];
+        $for c in range(TILE_TXCOLS):
+          $for j in range(4):
+            sums[i * TILE_TXCOLS * 4 + ${c} * 4 + ${j}] += accum * -8.0;
     }
 
     $if WEIGHT_STORAGE == "buffer":
@@ -109,13 +116,13 @@ void main() {
             packed_weight_tex = texelFetch(
               t_weight, ivec2(weight_txcol + ${c}, pos + r), 0);
 
-          qmat2_vec4 = (VEC4_T(packed_weight_tex >> 4) - 8.0);
+          qmat2_vec4 = VEC4_T(packed_weight_tex >> 4);
           qmat2[${c} * 4 * TILE_TXCOLS + 0] = qmat2_vec4.x;
           qmat2[${c} * 4 * TILE_TXCOLS + 1] = qmat2_vec4.y;
           qmat2[${c} * 4 * TILE_TXCOLS + 2] = qmat2_vec4.z;
           qmat2[${c} * 4 * TILE_TXCOLS + 3] = qmat2_vec4.w;
 
-          qmat2_vec4 = (VEC4_T(packed_weight_tex & 0x0F) - 8.0);
+          qmat2_vec4 = VEC4_T(packed_weight_tex & 0x0F);
           qmat2[${c} * 4 * TILE_TXCOLS + 4] = qmat2_vec4.x;
           qmat2[${c} * 4 * TILE_TXCOLS + 5] = qmat2_vec4.y;
           qmat2[${c} * 4 * TILE_TXCOLS + 6] = qmat2_vec4.z;